{"releases":[{"id":135924118,"tag":"v3.3.0","author":"hhaAndroid","name":"MMDetection v3.3.0 releases ","draft":false,"prerelease":false,"createdAt":"2024-01-05T06:10:48Z","publishedAt":"2024-01-05T06:24:15Z","markdown":"# MM Grounding DINO\r\n\r\n[An Open and Comprehensive Pipeline for Unified Object Grounding and Detection](https://arxiv.org/abs/2401.02361)\r\n\r\nGrounding-DINO is a state-of-the-art open-set detection model that tackles multiple vision tasks including Open-Vocabulary Detection (OVD), Phrase Grounding (PG), and Referring Expression Comprehension (REC). Its effectiveness has led to its widespread adoption as a mainstream architecture for various downstream applications. However, despite its significance, the original Grounding-DINO model lacks comprehensive public technical details due to the unavailability of its training code. To bridge this gap, we present MM-Grounding-DINO, an open-source, comprehensive, and user-friendly baseline, which is built with the MMDetection toolbox. It adopts abundant vision datasets for pre-training and various detection and grounding datasets for fine-tuning. We give a comprehensive analysis of each reported result and detailed settings for reproduction. The extensive experiments on the benchmarks mentioned demonstrate that our MM-Grounding-DINO-Tiny outperforms the Grounding-DINO-Tiny baseline. We release all our models to the research community.\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/4214e282-a553-4abf-b8a4-84ea566851c9\"/>\r\n</div>\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/fb14d1ee-5469-44d2-b865-aac9850c429c\"/>\r\n</div>\r\n\r\nDetail: https://github.com/open-mmlab/mmdetection/tree/main/configs/mm_grounding_dino","html":"<div class=\"markdown-heading\"><h1 class=\"heading-element\">MM Grounding DINO</h1><a id=\"user-content-mm-grounding-dino\" class=\"anchor\" aria-label=\"Permalink: MM Grounding DINO\" href=\"#mm-grounding-dino\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p><a href=\"https://arxiv.org/abs/2401.02361\" rel=\"nofollow\">An Open and Comprehensive Pipeline for Unified Object Grounding and Detection</a></p>\n<p>Grounding-DINO is a state-of-the-art open-set detection model that tackles multiple vision tasks including Open-Vocabulary Detection (OVD), Phrase Grounding (PG), and Referring Expression Comprehension (REC). Its effectiveness has led to its widespread adoption as a mainstream architecture for various downstream applications. However, despite its significance, the original Grounding-DINO model lacks comprehensive public technical details due to the unavailability of its training code. To bridge this gap, we present MM-Grounding-DINO, an open-source, comprehensive, and user-friendly baseline, which is built with the MMDetection toolbox. It adopts abundant vision datasets for pre-training and various detection and grounding datasets for fine-tuning. We give a comprehensive analysis of each reported result and detailed settings for reproduction. The extensive experiments on the benchmarks mentioned demonstrate that our MM-Grounding-DINO-Tiny outperforms the Grounding-DINO-Tiny baseline. We release all our models to the research community.</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/4214e282-a553-4abf-b8a4-84ea566851c9\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/4214e282-a553-4abf-b8a4-84ea566851c9\" style=\"max-width: 100%;\"></a>\n</div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/fb14d1ee-5469-44d2-b865-aac9850c429c\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/fb14d1ee-5469-44d2-b865-aac9850c429c\" style=\"max-width: 100%;\"></a>\n</div>\n<p>Detail: <a href=\"https://github.com/open-mmlab/mmdetection/tree/main/configs/mm_grounding_dino\">https://github.com/open-mmlab/mmdetection/tree/main/configs/mm_grounding_dino</a></p>\n","assets":[]},{"id":124753645,"tag":"v3.2.0","author":"hhaAndroid","name":"MMDetection v3.2.0 Release","draft":false,"prerelease":false,"createdAt":"2023-10-12T07:43:04Z","publishedAt":"2023-10-12T07:45:00Z","markdown":"# Highlight\r\n\r\n**v3.2.0** was released in 12/10/2023:\r\n\r\n**1. Detection Transformer SOTA Model Collection**\r\n(1) Supported four updated and stronger SOTA Transformer models: [DDQ](configs/ddq/README.md), [CO-DETR](projects/CO-DETR/README.md), [AlignDETR](projects/AlignDETR/README.md), and [H-DINO](projects/HDINO/README.md).\r\n(2) Based on CO-DETR, MMDet released a model with a COCO performance of 64.1 mAP.\r\n(3) Algorithms such as DINO support `AMP/Checkpoint/FrozenBN`, which can effectively reduce memory usage.\r\n\r\n**2. [Comprehensive Performance Comparison between CNN and Transformer](<(projects/RF100-Benchmark/README.md)>)**\r\nRF100 consists of a dataset collection of 100 real-world datasets, including 7 domains. It can be used to assess the performance differences of Transformer models like DINO and CNN-based algorithms under different scenarios and data volumes. Users can utilize this benchmark to quickly evaluate the robustness of their algorithms in various scenarios.\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\"/>\r\n</div>\r\n\r\n**3. Support for [GLIP](configs/glip/README.md) and [Grounding DINO](configs/grounding_dino/README.md) fine-tuning, the only algorithm library that supports Grounding DINO fine-tuning**\r\nThe Grounding DINO algorithm in MMDet is the only library that supports fine-tuning. Its performance is one point higher than the official version, and of course, GLIP also outperforms the official version.\r\nWe also provide a detailed process for training and evaluating Grounding DINO on custom datasets. Everyone is welcome to give it a try.\r\n\r\n|       Model        | Backbone |   Style   |  COCO mAP  | Official COCO mAP |\r\n| :----------------: | :------: | :-------: | :--------: | :---------------: |\r\n|  Grounding DINO-T  |  Swin-T  | Zero-shot |    48.5    |       48.4        |\r\n|  Grounding DINO-T  |  Swin-T  | Finetune  | 58.1(+0.9) |       57.2        |\r\n|  Grounding DINO-B  |  Swin-B  | Zero-shot |    56.9    |       56.7        |\r\n|  Grounding DINO-B  |  Swin-B  | Finetune  |    59.7    |                   |\r\n| Grounding DINO-R50 |   R50    |  Scratch  | 48.9(+0.8) |       48.1        |\r\n\r\n**4. Support for the open-vocabulary detection algorithm [Detic](projects/Detic_new/README.md) and multi-dataset joint training.**\r\n**5. Training detection models using [FSDP and DeepSpeed](<(projects/example_largemodel/README.md)>).**\r\n\r\n| ID  | AMP | GC of Backbone | GC of Encoder | FSDP | Peak Mem (GB) | Iter Time (s) |\r\n| :-: | :-: | :------------: | :-----------: | :--: | :-----------: | :-----------: |\r\n|  1  |     |                |               |      |   49 (A100)   |      0.9      |\r\n|  2  |  √  |                |               |      |   39 (A100)   |      1.2      |\r\n|  3  |     |       √        |               |      |   33 (A100)   |      1.1      |\r\n|  4  |  √  |       √        |               |      |   25 (A100)   |      1.3      |\r\n|  5  |     |       √        |       √       |      |      18       |      2.2      |\r\n|  6  |  √  |       √        |       √       |      |      13       |      1.6      |\r\n|  7  |     |       √        |       √       |  √   |      14       |      2.9      |\r\n|  8  |  √  |       √        |       √       |  √   |      8.5      |      2.4      |\r\n\r\n**6. Support for the [V3Det](configs/v3det/README.md) dataset, a large-scale detection dataset with over 13,000 categories.**\r\n\r\n<div align=center>\r\n    <img width=960 src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\"/>\r\n</div>\r\n\r\n# 亮点\r\n\r\n**v3.2.0** 版本已经在 2023.10.12 发布：\r\n\r\n**1. 检测 Transformer SOTA 模型大合集**\r\n(1) 支持了 [DDQ](configs/ddq/README.md)、[CO-DETR](projects/CO-DETR/README.md)、[AlignDETR](projects/AlignDETR/README.md) 和 [H-DINO](projects/HDINO/README.md) 4 个更新更强的 SOTA Transformer 模型\r\n(2) 基于 CO-DETR,  MMDet 中发布了 COCO 性能为 64.1 mAP 的模型\r\n(3) DINO 等算法支持 AMP/Checkpoint/FrozenBN，可以有效降低显存\r\n\r\n**2. [提供了全面的 CNN 和 Transformer 的性能对比](projects/RF100-Benchmark/README_zh-CN.md)**\r\nRF100 是由 100 个现实收集的数据集组成，包括 7 个域，可以验证 DINO 等 Transformer 模型和 CNN 类算法在不同场景不同数据量下的性能差异。用户可以用这个 Benchmark 快速验证自己的算法在不同场景下的鲁棒性。\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\"/>\r\n</div>\r\n\r\n**3. 支持了 [GLIP](configs/glip/README.md) 和 [Grounding DINO](configs/grounding_dino/README.md) 微调，全网唯一支持 Grounding DINO 微调**\r\nMMDet 中的 Grounding DINO 是全网唯一支持微调的算法库，且性能高于官方 1 个点，当然 GLIP 也比官方高。\r\n我们还提供了详细的 Grounding DINO 在自定义数据集上训练评估的流程，欢迎大家试用。\r\n\r\n|       Model        | Backbone |   Style   |  COCO mAP  | Official COCO mAP |\r\n| :----------------: | :------: | :-------: | :--------: | :---------------: |\r\n|  Grounding DINO-T  |  Swin-T  | Zero-shot |    48.5    |       48.4        |\r\n|  Grounding DINO-T  |  Swin-T  | Finetune  | 58.1(+0.9) |       57.2        |\r\n|  Grounding DINO-B  |  Swin-B  | Zero-shot |    56.9    |       56.7        |\r\n|  Grounding DINO-B  |  Swin-B  | Finetune  |    59.7    |                   |\r\n| Grounding DINO-R50 |   R50    |  Scratch  | 48.9(+0.8) |       48.1        |\r\n\r\n**4. 支持开放词汇检测算法 [Detic](projects/Detic_new/README.md) 并提供多数据集联合训练可能**\r\n\r\n**5. 轻松使用 [FSDP 和 DeepSpeed 训练检测模型](projects/example_largemodel/README_zh-CN.md)**\r\n\r\n| ID  | AMP | GC of Backbone | GC of Encoder | FSDP | Peak Mem (GB) | Iter Time (s) |\r\n| :-: | :-: | :------------: | :-----------: | :--: | :-----------: | :-----------: |\r\n|  1  |     |                |               |      |   49 (A100)   |      0.9      |\r\n|  2  |  √  |                |               |      |   39 (A100)   |      1.2      |\r\n|  3  |     |       √        |               |      |   33 (A100)   |      1.1      |\r\n|  4  |  √  |       √        |               |      |   25 (A100)   |      1.3      |\r\n|  5  |     |       √        |       √       |      |      18       |      2.2      |\r\n|  6  |  √  |       √        |       √       |      |      13       |      1.6      |\r\n|  7  |     |       √        |       √       |  √   |      14       |      2.9      |\r\n|  8  |  √  |       √        |       √       |  √   |      8.5      |      2.4      |\r\n\r\n**6. 支持了 [V3Det](configs/v3det/README.md) 1.3w+ 类别的超大词汇检测数据集**\r\n\r\n<div align=center>\r\n    <img width=960 src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\"/>\r\n</div>\r\n","html":"<div class=\"markdown-heading\"><h1 class=\"heading-element\">Highlight</h1><a id=\"user-content-highlight\" class=\"anchor\" aria-label=\"Permalink: Highlight\" href=\"#highlight\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p><strong>v3.2.0</strong> was released in 12/10/2023:</p>\n<p><strong>1. Detection Transformer SOTA Model Collection</strong>\n(1) Supported four updated and stronger SOTA Transformer models: <a href=\"configs/ddq/README.md\">DDQ</a>, <a href=\"projects/CO-DETR/README.md\">CO-DETR</a>, <a href=\"projects/AlignDETR/README.md\">AlignDETR</a>, and <a href=\"projects/HDINO/README.md\">H-DINO</a>.\n(2) Based on CO-DETR, MMDet released a model with a COCO performance of 64.1 mAP.\n(3) Algorithms such as DINO support <code>AMP/Checkpoint/FrozenBN</code>, which can effectively reduce memory usage.</p>\n<p><strong>2. <a href=\"(projects/RF100-Benchmark/README.md)\">Comprehensive Performance Comparison between CNN and Transformer</a></strong>\nRF100 consists of a dataset collection of 100 real-world datasets, including 7 domains. It can be used to assess the performance differences of Transformer models like DINO and CNN-based algorithms under different scenarios and data volumes. Users can utilize this benchmark to quickly evaluate the robustness of their algorithms in various scenarios.</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>3. Support for <a href=\"configs/glip/README.md\">GLIP</a> and <a href=\"configs/grounding_dino/README.md\">Grounding DINO</a> fine-tuning, the only algorithm library that supports Grounding DINO fine-tuning</strong>\nThe Grounding DINO algorithm in MMDet is the only library that supports fine-tuning. Its performance is one point higher than the official version, and of course, GLIP also outperforms the official version.\nWe also provide a detailed process for training and evaluating Grounding DINO on custom datasets. Everyone is welcome to give it a try.</p>\n<table>\n<thead>\n<tr>\n<th align=\"center\">Model</th>\n<th align=\"center\">Backbone</th>\n<th align=\"center\">Style</th>\n<th align=\"center\">COCO mAP</th>\n<th align=\"center\">Official COCO mAP</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"center\">Grounding DINO-T</td>\n<td align=\"center\">Swin-T</td>\n<td align=\"center\">Zero-shot</td>\n<td align=\"center\">48.5</td>\n<td align=\"center\">48.4</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-T</td>\n<td align=\"center\">Swin-T</td>\n<td align=\"center\">Finetune</td>\n<td align=\"center\">58.1(+0.9)</td>\n<td align=\"center\">57.2</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-B</td>\n<td align=\"center\">Swin-B</td>\n<td align=\"center\">Zero-shot</td>\n<td align=\"center\">56.9</td>\n<td align=\"center\">56.7</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-B</td>\n<td align=\"center\">Swin-B</td>\n<td align=\"center\">Finetune</td>\n<td align=\"center\">59.7</td>\n<td align=\"center\"></td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-R50</td>\n<td align=\"center\">R50</td>\n<td align=\"center\">Scratch</td>\n<td align=\"center\">48.9(+0.8)</td>\n<td align=\"center\">48.1</td>\n</tr>\n</tbody>\n</table>\n<p><strong>4. Support for the open-vocabulary detection algorithm <a href=\"projects/Detic_new/README.md\">Detic</a> and multi-dataset joint training.</strong>\n<strong>5. Training detection models using <a href=\"(projects/example_largemodel/README.md)\">FSDP and DeepSpeed</a>.</strong></p>\n<table>\n<thead>\n<tr>\n<th align=\"center\">ID</th>\n<th align=\"center\">AMP</th>\n<th align=\"center\">GC of Backbone</th>\n<th align=\"center\">GC of Encoder</th>\n<th align=\"center\">FSDP</th>\n<th align=\"center\">Peak Mem (GB)</th>\n<th align=\"center\">Iter Time (s)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"center\">1</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">49 (A100)</td>\n<td align=\"center\">0.9</td>\n</tr>\n<tr>\n<td align=\"center\">2</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">39 (A100)</td>\n<td align=\"center\">1.2</td>\n</tr>\n<tr>\n<td align=\"center\">3</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">33 (A100)</td>\n<td align=\"center\">1.1</td>\n</tr>\n<tr>\n<td align=\"center\">4</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">25 (A100)</td>\n<td align=\"center\">1.3</td>\n</tr>\n<tr>\n<td align=\"center\">5</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\">18</td>\n<td align=\"center\">2.2</td>\n</tr>\n<tr>\n<td align=\"center\">6</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\">13</td>\n<td align=\"center\">1.6</td>\n</tr>\n<tr>\n<td align=\"center\">7</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">14</td>\n<td align=\"center\">2.9</td>\n</tr>\n<tr>\n<td align=\"center\">8</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">8.5</td>\n<td align=\"center\">2.4</td>\n</tr>\n</tbody>\n</table>\n<p><strong>6. Support for the <a href=\"configs/v3det/README.md\">V3Det</a> dataset, a large-scale detection dataset with over 13,000 categories.</strong></p>\n<div align=\"center\">\n    <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\"><img width=\"960\" src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\" style=\"max-width: 100%;\"></a>\n</div>\n<div class=\"markdown-heading\"><h1 class=\"heading-element\">亮点</h1><a id=\"user-content-亮点\" class=\"anchor\" aria-label=\"Permalink: 亮点\" href=\"#亮点\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p><strong>v3.2.0</strong> 版本已经在 2023.10.12 发布：</p>\n<p><strong>1. 检测 Transformer SOTA 模型大合集</strong>\n(1) 支持了 <a href=\"configs/ddq/README.md\">DDQ</a>、<a href=\"projects/CO-DETR/README.md\">CO-DETR</a>、<a href=\"projects/AlignDETR/README.md\">AlignDETR</a> 和 <a href=\"projects/HDINO/README.md\">H-DINO</a> 4 个更新更强的 SOTA Transformer 模型\n(2) 基于 CO-DETR,  MMDet 中发布了 COCO 性能为 64.1 mAP 的模型\n(3) DINO 等算法支持 AMP/Checkpoint/FrozenBN，可以有效降低显存</p>\n<p><strong>2. <a href=\"projects/RF100-Benchmark/README_zh-CN.md\">提供了全面的 CNN 和 Transformer 的性能对比</a></strong>\nRF100 是由 100 个现实收集的数据集组成，包括 7 个域，可以验证 DINO 等 Transformer 模型和 CNN 类算法在不同场景不同数据量下的性能差异。用户可以用这个 Benchmark 快速验证自己的算法在不同场景下的鲁棒性。</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/86420903-36a8-410d-9251-4304b9704f7d\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>3. 支持了 <a href=\"configs/glip/README.md\">GLIP</a> 和 <a href=\"configs/grounding_dino/README.md\">Grounding DINO</a> 微调，全网唯一支持 Grounding DINO 微调</strong>\nMMDet 中的 Grounding DINO 是全网唯一支持微调的算法库，且性能高于官方 1 个点，当然 GLIP 也比官方高。\n我们还提供了详细的 Grounding DINO 在自定义数据集上训练评估的流程，欢迎大家试用。</p>\n<table>\n<thead>\n<tr>\n<th align=\"center\">Model</th>\n<th align=\"center\">Backbone</th>\n<th align=\"center\">Style</th>\n<th align=\"center\">COCO mAP</th>\n<th align=\"center\">Official COCO mAP</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"center\">Grounding DINO-T</td>\n<td align=\"center\">Swin-T</td>\n<td align=\"center\">Zero-shot</td>\n<td align=\"center\">48.5</td>\n<td align=\"center\">48.4</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-T</td>\n<td align=\"center\">Swin-T</td>\n<td align=\"center\">Finetune</td>\n<td align=\"center\">58.1(+0.9)</td>\n<td align=\"center\">57.2</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-B</td>\n<td align=\"center\">Swin-B</td>\n<td align=\"center\">Zero-shot</td>\n<td align=\"center\">56.9</td>\n<td align=\"center\">56.7</td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-B</td>\n<td align=\"center\">Swin-B</td>\n<td align=\"center\">Finetune</td>\n<td align=\"center\">59.7</td>\n<td align=\"center\"></td>\n</tr>\n<tr>\n<td align=\"center\">Grounding DINO-R50</td>\n<td align=\"center\">R50</td>\n<td align=\"center\">Scratch</td>\n<td align=\"center\">48.9(+0.8)</td>\n<td align=\"center\">48.1</td>\n</tr>\n</tbody>\n</table>\n<p><strong>4. 支持开放词汇检测算法 <a href=\"projects/Detic_new/README.md\">Detic</a> 并提供多数据集联合训练可能</strong></p>\n<p><strong>5. 轻松使用 <a href=\"projects/example_largemodel/README_zh-CN.md\">FSDP 和 DeepSpeed 训练检测模型</a></strong></p>\n<table>\n<thead>\n<tr>\n<th align=\"center\">ID</th>\n<th align=\"center\">AMP</th>\n<th align=\"center\">GC of Backbone</th>\n<th align=\"center\">GC of Encoder</th>\n<th align=\"center\">FSDP</th>\n<th align=\"center\">Peak Mem (GB)</th>\n<th align=\"center\">Iter Time (s)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"center\">1</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">49 (A100)</td>\n<td align=\"center\">0.9</td>\n</tr>\n<tr>\n<td align=\"center\">2</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">39 (A100)</td>\n<td align=\"center\">1.2</td>\n</tr>\n<tr>\n<td align=\"center\">3</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">33 (A100)</td>\n<td align=\"center\">1.1</td>\n</tr>\n<tr>\n<td align=\"center\">4</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\"></td>\n<td align=\"center\">25 (A100)</td>\n<td align=\"center\">1.3</td>\n</tr>\n<tr>\n<td align=\"center\">5</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\">18</td>\n<td align=\"center\">2.2</td>\n</tr>\n<tr>\n<td align=\"center\">6</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\"></td>\n<td align=\"center\">13</td>\n<td align=\"center\">1.6</td>\n</tr>\n<tr>\n<td align=\"center\">7</td>\n<td align=\"center\"></td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">14</td>\n<td align=\"center\">2.9</td>\n</tr>\n<tr>\n<td align=\"center\">8</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">√</td>\n<td align=\"center\">8.5</td>\n<td align=\"center\">2.4</td>\n</tr>\n</tbody>\n</table>\n<p><strong>6. 支持了 <a href=\"configs/v3det/README.md\">V3Det</a> 1.3w+ 类别的超大词汇检测数据集</strong></p>\n<div align=\"center\">\n    <a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\"><img width=\"960\" src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9c216387-02be-46e6-b0f2-b856f80f6d84\" style=\"max-width: 100%;\"></a>\n</div>\n","assets":[]},{"id":110329337,"tag":"v3.1.0","author":"hhaAndroid","name":"MMDetection  v3.1.0 Release","draft":false,"prerelease":false,"createdAt":"2023-06-30T07:37:44Z","publishedAt":"2023-06-30T07:41:30Z","markdown":"# Highlights\r\n\r\n- Supports tracking algorithms including multi-object tracking (MOT) algorithms SORT, DeepSORT, StrongSORT, OCSORT, ByteTrack, QDTrack, and video instance segmentation (VIS) algorithm MaskTrackRCNN, Mask2Former-VIS.\r\n- Support [ViTDet](https://github.com/open-mmlab/mmdetection/tree/dev-3.x/projects/ViTDet)\r\n- Supports inference and evaluation of multimodal algorithms [GLIP](https://github.com/open-mmlab/mmdetection/tree/dev-3.x/configs/glip) and [XDecoder](https://github.com/open-mmlab/mmdetection/tree/dev-3.x/projects/XDecoder), and also supports datasets such as COCO semantic segmentation, COCO Caption, ADE20k general segmentation, and RefCOCO. GLIP fine-tuning will be supported in the future.\r\n- Provides a [gradio demo](https://github.com/open-mmlab/mmdetection/blob/dev-3.x/projects/gradio_demo/README.md) for image type tasks of MMDetection, making it easy for users to experience.\r\n\r\n# Exciting Features\r\n\r\n## GLIP inference and evaluation\r\n\r\ns multimodal vision algorithms continue to evolve, MMDetection has also supported such algorithms. This section demonstrates how to use the demo and eval scripts corresponding to multimodal algorithms using the GLIP algorithm and model as the example. Moreover, MMDetection integrated a [gradio_demo project](../../../projects/gradio_demo/), which allows developers to quickly play with all image input tasks in MMDetection on their local devices. Check the [document](../../../projects/gradio_demo/README.md) for more details.\r\n\r\n### Preparation\r\n\r\nPlease first make sure that you have the correct dependencies installed:\r\n\r\n```shell\r\n# if source\r\npip install -r requirements/multimodal.txt\r\n\r\n# if wheel\r\nmim install mmdet[multimodal]\r\n```\r\n\r\nMMDetection has already implemented GLIP algorithms and provided the weights, you can download directly from urls:\r\n\r\n```shell\r\ncd mmdetection\r\nwget https://download.openmmlab.com/mmdetection/v3.0/glip/glip_tiny_a_mmdet-b3654169.pth\r\n```\r\n\r\n### Inference\r\n\r\nOnce the model is successfully downloaded, you can use the `demo/image_demo.py` script to run the inference.\r\n\r\n```shell\r\npython demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts bench\r\n```\r\n\r\nDemo result will be similar to this:\r\n\r\n<div align=center>\r\n<img src=\"https://user-images.githubusercontent.com/17425982/234547841-266476c8-f987-4832-8642-34357be621c6.png\" height=\"300\"/>\r\n</div>\r\n\r\nIf users would like to detect multiple targets, please declare them in the format of `xx . xx .` after the `--texts`.\r\n\r\n```shell\r\npython demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts 'bench . car .'\r\n```\r\n\r\nAnd the result will be like this one:\r\n\r\n<div align=center>\r\n<img src=\"https://user-images.githubusercontent.com/17425982/234548156-ef9bbc2e-7605-4867-abe6-048b8578893d.png\" height=\"300\"/>\r\n</div>\r\n\r\nYou can also use a sentence as the input prompt for the `--texts` field, for example:\r\n\r\n```shell\r\npython demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts 'There are a lot of cars here.'\r\n```\r\n\r\nThe result will be similar to this:\r\n\r\n<div align=center>\r\n<img src=\"https://user-images.githubusercontent.com/17425982/234548490-d2e0a16d-1aad-4708-aea0-c829634fabbd.png\" height=\"300\"/>\r\n</div>\r\n\r\n### Evaluation\r\n\r\nThe GLIP implementation in MMDetection does not have any performance degradation, our benchmark is as follows:\r\n\r\n| Model                   | official mAP | mmdet mAP |\r\n| ----------------------- | :----------: | :-------: |\r\n| glip_A_Swin_T_O365.yaml |     42.9     |   43.0    |\r\n| glip_Swin_T_O365.yaml   |     44.9     |   44.9    |\r\n| glip_Swin_L.yaml        |     51.4     |   51.3    |\r\n\r\nUsers can use the test script we provided to run evaluation as well. Here is a basic example:\r\n\r\n```shell\r\n# 1 gpu\r\npython tools/test.py configs/glip/glip_atss_swin-t_fpn_dyhead_pretrain_obj365.py glip_tiny_a_mmdet-b3654169.pth\r\n\r\n# 8 GPU\r\n./tools/dist_test.sh configs/glip/glip_atss_swin-t_fpn_dyhead_pretrain_obj365.py glip_tiny_a_mmdet-b3654169.pth 8\r\n```\r\n\r\nThe result will be similar to this:\r\n\r\n```shell\r\nAverage Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.428\r\nAverage Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=1000 ] = 0.594\r\nAverage Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=1000 ] = 0.466\r\nAverage Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.300\r\nAverage Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.477\r\nAverage Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.534\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.634\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 ] = 0.634\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=1000 ] = 0.634\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.473\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.690\r\nAverage Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.789\r\n```\r\n\r\n## XDecoder\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/cb126615-9402-4c19-8ea9-133722d7519c\" width=\"70%\"/>\r\n</div>\r\n\r\n## Installation\r\n\r\n```shell\r\n# if source\r\npip install -r requirements/multimodal.txt\r\n\r\n# if wheel\r\nmim install mmdet[multimodal]\r\n```\r\n\r\n## How to use it?\r\n\r\nFor convenience, you can download the weights to the `mmdetection` root dir\r\n\r\n```shell\r\nwget https://download.openmmlab.com/mmdetection/v3.0/xdecoder/xdecoder_focalt_last_novg.pt\r\nwget https://download.openmmlab.com/mmdetection/v3.0/xdecoder/xdecoder_focalt_best_openseg.pt\r\n```\r\n\r\nThe above two weights are directly copied from the official website without any modification. The specific source is https://github.com/microsoft/X-Decoder\r\n\r\nFor convenience of demonstration, please download [the folder](https://github.com/microsoft/X-Decoder/tree/main/images) and place it in the root directory of mmdetection.\r\n\r\n**(1) Open Vocabulary Semantic Segmentation**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/animals.png configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py --weights ../../xdecoder_focalt_last_novg.pt --texts zebra.giraffe\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/c397c0ed-859a-4004-8725-78a591742bc8\" width=\"70%\"/>\r\n</div>\r\n\r\n**(2) Open Vocabulary Instance Segmentation**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/owls.jpeg configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py --weights ../../xdecoder_focalt_last_novg.pt --texts owl\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/494b0b1c-4a42-4019-97ae-d33ee68af3d2\" width=\"70%\"/>\r\n</div>\r\n\r\n**(3) Open Vocabulary Panoptic Segmentation**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/street.jpg configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py --weights ../../xdecoder_focalt_last_novg.pt  --text car.person --stuff-text tree.sky\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9ad1e0f4-75ce-4e37-a5cc-83e0e8a722ed\" width=\"70%\"/>\r\n</div>\r\n\r\n**(4) Referring Expression Segmentation**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/fruit.jpg configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py --weights ../../xdecoder_focalt_last_novg.pt  --text \"The larger watermelon. The front white flower. White tea pot.\"\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/f3ecdb50-20f0-4dc4-aa9c-90995ae04893\" width=\"70%\"/>\r\n</div>\r\n\r\n**(5) Image Caption**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/penguin.jpeg configs/xdecoder-tiny_zeroshot_caption_coco2014.py --weights ../../xdecoder_focalt_last_novg.pt\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/7690ab79-791e-4011-ab0c-01f46c4a3d80\" width=\"70%\"/>\r\n</div>\r\n\r\n**(6) Referring Expression Image Caption**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/fruit.jpg configs/xdecoder-tiny_zeroshot_ref-caption.py --weights ../../xdecoder_focalt_last_novg.pt --text 'White tea pot'\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/bae2fdba-0172-4fc8-8ad1-73b54c64ec30\" width=\"70%\"/>\r\n</div>\r\n\r\n**(7) Text Image Region Retrieval**\r\n\r\n```shell\r\ncd projects/XDecoder\r\npython demo.py ../../images/coco configs/xdecoder-tiny_zeroshot_text-image-retrieval.py --weights ../../xdecoder_focalt_last_novg.pt --text 'pizza on the plate'\r\n```\r\n\r\n```text\r\nThe image that best matches the given text is ../../images/coco/000.jpg and probability is 0.998\r\n```\r\n\r\n<div align=center>\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/479de6b2-88e7-41f0-8228-4b9a48f52954\" width=\"70%\"/>\r\n</div>\r\n\r\nWe have also prepared a gradio program in the `projects/gradio_demo` directory, which you can run interactively all the inference supported by mmdetection in your browser.\r\n\r\n## Models and results\r\n\r\n### Semantic segmentation on ADE20K\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation).\r\n\r\n**Test Command**\r\n\r\nSince semantic segmentation is a pixel-level task, we don't need to use a threshold to filter out low-confidence predictions. So we set `model.test_cfg.use_thr_for_mc=False` in the test command.\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-semseg_ade20k.py xdecoder_focalt_best_openseg.pt 8 --cfg-options model.test_cfg.use_thr_for_mc=False\r\n```\r\n\r\n| Model                             | mIoU  | mIOU(official) |                                Config                                |\r\n| :-------------------------------- | :---: | :------------: | :------------------------------------------------------------------: |\r\n| `xdecoder_focalt_best_openseg.pt` | 25.24 |     25.13      | [config](configs/xdecoder-tiny_zeroshot_open-vocab-semseg_ade20k.py) |\r\n\r\n### Instance segmentation on ADE20K\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation).\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-instance_ade20k.py xdecoder_focalt_best_openseg.pt 8\r\n```\r\n\r\n| Model                             | mIoU | mIOU(official) |                                 Config                                 |\r\n| :-------------------------------- | :--: | :------------: | :--------------------------------------------------------------------: |\r\n| `xdecoder_focalt_best_openseg.pt` | 10.1 |      10.1      | [config](configs/xdecoder-tiny_zeroshot_open-vocab-instance_ade20k.py) |\r\n\r\n### Panoptic segmentation on ADE20K\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation).\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_ade20k.py xdecoder_focalt_best_openseg.pt 8\r\n```\r\n\r\n| Model                             | mIoU  | mIOU(official) |                                 Config                                 |\r\n| :-------------------------------- | :---: | :------------: | :--------------------------------------------------------------------: |\r\n| `xdecoder_focalt_best_openseg.pt` | 19.11 |     18.97      | [config](configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_ade20k.py) |\r\n\r\n### Semantic segmentation on COCO2017\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#coco-semantic-dataset-preparation) of `(2) use panoptic dataset` part.\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py xdecoder_focalt_last_novg.pt 8 --cfg-options model.test_cfg.use_thr_for_mc=False\r\n```\r\n\r\n| Model                                           | mIOU | mIOU(official) |                               Config                               |\r\n| :---------------------------------------------- | :--: | :------------: | :----------------------------------------------------------------: |\r\n| `xdecoder-tiny_zeroshot_open-vocab-semseg_coco` | 62.1 |      62.1      | [config](configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py) |\r\n\r\n### Instance segmentation on COCO2017\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#basic-detection-dataset-preparation).\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py xdecoder_focalt_last_novg.pt 8\r\n```\r\n\r\n| Model                                             | Mask mAP | Mask mAP(official) |                                Config                                |\r\n| :------------------------------------------------ | :------: | :----------------: | :------------------------------------------------------------------: |\r\n| `xdecoder-tiny_zeroshot_open-vocab-instance_coco` |   39.8   |        39.7        | [config](configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py) |\r\n\r\n### Panoptic segmentation on COCO2017\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#basic-detection-dataset-preparation).\r\n\r\n```shell\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py xdecoder_focalt_last_novg.pt 8\r\n```\r\n\r\n| Model                                             |  PQ   | PQ(official) |                                Config                                |\r\n| :------------------------------------------------ | :---: | :----------: | :------------------------------------------------------------------: |\r\n| `xdecoder-tiny_zeroshot_open-vocab-panoptic_coco` | 51.42 |    51.16     | [config](configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py) |\r\n\r\n### Referring segmentation on RefCOCO\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#refcoco-dataset-preparation).\r\n\r\n```shell\r\n./tools/dist_test.sh  projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py xdecoder_focalt_last_novg.pt 8  --cfg-options test_dataloader.dataset.split='val'\r\n```\r\n\r\n| Model                          |  text mode   |  cIoU   | cIOU(official) |                                 Config                                  |\r\n| :----------------------------- | :----------: | :-----: | :------------: | :---------------------------------------------------------------------: |\r\n| `xdecoder_focalt_last_novg.pt` | select first | 58.8415 |     57.85      | [config](configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py) |\r\n| `xdecoder_focalt_last_novg.pt` |   original   | 60.0321 |       -        | [config](configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py) |\r\n| `xdecoder_focalt_last_novg.pt` |    concat    | 60.3551 |       -        | [config](configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py) |\r\n\r\n**Note:**\r\n\r\n1. If you set the scale of `Resize` to (1024, 512), the result will be `57.69`.\r\n2. `text mode` is the `RefCoCoDataset` parameter in MMDetection, it determines the texts loaded to the data list. It can be set to `select_first`, `original`, `concat` and `random`.\r\n   - `select_first`: select the first text in the text list as the description to an instance.\r\n   - `original`: use all texts in the text list as the description to an instance.\r\n   - `concat`: concatenate all texts in the text list as the description to an instance.\r\n   - `random`: randomly select one text in the text list as the description to an instance, usually used for training.\r\n\r\n### Image Caption on COCO2014\r\n\r\nPrepare your dataset according to the [docs](../../docs/en/user_guides/dataset_prepare.md#coco-caption-dataset-preparation).\r\n\r\nBefore testing, you need to install jdk 1.8, otherwise it will prompt that java does not exist during the evaluation process\r\n\r\n```\r\n./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_caption_coco2014.py xdecoder_focalt_last_novg.pt 8\r\n```\r\n\r\n| Model                                     | BLEU-4 | CIDER  |                            Config                            |\r\n| :---------------------------------------- | :----: | :----: | :----------------------------------------------------------: |\r\n| `xdecoder-tiny_zeroshot_caption_coco2014` | 35.26  | 116.81 | [config](configs/xdecoder-tiny_zeroshot_caption_coco2014.py) |\r\n\r\n\r\n## Gradio Demo\r\n\r\n<div align=\"center\">\r\n<img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/6c29886f-ae7a-4a55-8be4-352ee85b7d3e\"/>\r\n</div>\r\n\r\nPlease refer to https://github.com/open-mmlab/mmdetection/blob/dev-3.x/projects/gradio_demo/README.md for details.\r\n\r\n## Contributors\r\n\r\nA total of 30 developers contributed to this release.\r\n\r\nThanks @jjjkkkjjj @lovelykite, @minato-ellie, @freepoet, @wufan-tb, @yalibian, @keyakiluo, @gihanjayatilaka, @i-aki-y, @xin-li-67, @RangeKing, @JingweiZhang12, @MambaWong, @lucianovk, @tall-josh, @xiuqhou, @jamiechoi1995, @YQisme, @yechenzhi, @bjzhb666, @xiexinch, @jamiechoi1995, @yarkable, @Renzhihan, @nijkah, @amaizr, @Lum1104, @zwhus, @Czm369, @hhaAndroid","html":"<div class=\"markdown-heading\"><h1 class=\"heading-element\">Highlights</h1><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Supports tracking algorithms including multi-object tracking (MOT) algorithms SORT, DeepSORT, StrongSORT, OCSORT, ByteTrack, QDTrack, and video instance segmentation (VIS) algorithm MaskTrackRCNN, Mask2Former-VIS.</li>\n<li>Support <a href=\"https://github.com/open-mmlab/mmdetection/tree/dev-3.x/projects/ViTDet\">ViTDet</a>\n</li>\n<li>Supports inference and evaluation of multimodal algorithms <a href=\"https://github.com/open-mmlab/mmdetection/tree/dev-3.x/configs/glip\">GLIP</a> and <a href=\"https://github.com/open-mmlab/mmdetection/tree/dev-3.x/projects/XDecoder\">XDecoder</a>, and also supports datasets such as COCO semantic segmentation, COCO Caption, ADE20k general segmentation, and RefCOCO. GLIP fine-tuning will be supported in the future.</li>\n<li>Provides a <a href=\"https://github.com/open-mmlab/mmdetection/blob/dev-3.x/projects/gradio_demo/README.md\">gradio demo</a> for image type tasks of MMDetection, making it easy for users to experience.</li>\n</ul>\n<div class=\"markdown-heading\"><h1 class=\"heading-element\">Exciting Features</h1><a id=\"user-content-exciting-features\" class=\"anchor\" aria-label=\"Permalink: Exciting Features\" href=\"#exciting-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">GLIP inference and evaluation</h2><a id=\"user-content-glip-inference-and-evaluation\" class=\"anchor\" aria-label=\"Permalink: GLIP inference and evaluation\" href=\"#glip-inference-and-evaluation\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>s multimodal vision algorithms continue to evolve, MMDetection has also supported such algorithms. This section demonstrates how to use the demo and eval scripts corresponding to multimodal algorithms using the GLIP algorithm and model as the example. Moreover, MMDetection integrated a <a href=\"../../../projects/gradio_demo/\">gradio_demo project</a>, which allows developers to quickly play with all image input tasks in MMDetection on their local devices. Check the <a href=\"../../../projects/gradio_demo/README.md\">document</a> for more details.</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Preparation</h3><a id=\"user-content-preparation\" class=\"anchor\" aria-label=\"Permalink: Preparation\" href=\"#preparation\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Please first make sure that you have the correct dependencies installed:</p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c\"><span class=\"pl-c\">#</span> if source</span>\npip install -r requirements/multimodal.txt\n\n<span class=\"pl-c\"><span class=\"pl-c\">#</span> if wheel</span>\nmim install mmdet[multimodal]</pre></div>\n<p>MMDetection has already implemented GLIP algorithms and provided the weights, you can download directly from urls:</p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> mmdetection\nwget https://download.openmmlab.com/mmdetection/v3.0/glip/glip_tiny_a_mmdet-b3654169.pth</pre></div>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Inference</h3><a id=\"user-content-inference\" class=\"anchor\" aria-label=\"Permalink: Inference\" href=\"#inference\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Once the model is successfully downloaded, you can use the <code>demo/image_demo.py</code> script to run the inference.</p>\n<div class=\"highlight highlight-source-shell\"><pre>python demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts bench</pre></div>\n<p>Demo result will be similar to this:</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https://user-images.githubusercontent.com/17425982/234547841-266476c8-f987-4832-8642-34357be621c6.png\"><img src=\"https://user-images.githubusercontent.com/17425982/234547841-266476c8-f987-4832-8642-34357be621c6.png\" height=\"300\" style=\"max-width: 100%; height: auto; max-height: 300px;\"></a>\n</div>\n<p>If users would like to detect multiple targets, please declare them in the format of <code>xx . xx .</code> after the <code>--texts</code>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>python demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts <span class=\"pl-s\"><span class=\"pl-pds\">'</span>bench . car .<span class=\"pl-pds\">'</span></span></pre></div>\n<p>And the result will be like this one:</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https://user-images.githubusercontent.com/17425982/234548156-ef9bbc2e-7605-4867-abe6-048b8578893d.png\"><img src=\"https://user-images.githubusercontent.com/17425982/234548156-ef9bbc2e-7605-4867-abe6-048b8578893d.png\" height=\"300\" style=\"max-width: 100%; height: auto; max-height: 300px;\"></a>\n</div>\n<p>You can also use a sentence as the input prompt for the <code>--texts</code> field, for example:</p>\n<div class=\"highlight highlight-source-shell\"><pre>python demo/image_demo.py demo/demo.jpg glip_tiny_a_mmdet-b3654169.pth --texts <span class=\"pl-s\"><span class=\"pl-pds\">'</span>There are a lot of cars here.<span class=\"pl-pds\">'</span></span></pre></div>\n<p>The result will be similar to this:</p>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https://user-images.githubusercontent.com/17425982/234548490-d2e0a16d-1aad-4708-aea0-c829634fabbd.png\"><img src=\"https://user-images.githubusercontent.com/17425982/234548490-d2e0a16d-1aad-4708-aea0-c829634fabbd.png\" height=\"300\" style=\"max-width: 100%; height: auto; max-height: 300px;\"></a>\n</div>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Evaluation</h3><a id=\"user-content-evaluation\" class=\"anchor\" aria-label=\"Permalink: Evaluation\" href=\"#evaluation\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The GLIP implementation in MMDetection does not have any performance degradation, our benchmark is as follows:</p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th align=\"center\">official mAP</th>\n<th align=\"center\">mmdet mAP</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>glip_A_Swin_T_O365.yaml</td>\n<td align=\"center\">42.9</td>\n<td align=\"center\">43.0</td>\n</tr>\n<tr>\n<td>glip_Swin_T_O365.yaml</td>\n<td align=\"center\">44.9</td>\n<td align=\"center\">44.9</td>\n</tr>\n<tr>\n<td>glip_Swin_L.yaml</td>\n<td align=\"center\">51.4</td>\n<td align=\"center\">51.3</td>\n</tr>\n</tbody>\n</table>\n<p>Users can use the test script we provided to run evaluation as well. Here is a basic example:</p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c\"><span class=\"pl-c\">#</span> 1 gpu</span>\npython tools/test.py configs/glip/glip_atss_swin-t_fpn_dyhead_pretrain_obj365.py glip_tiny_a_mmdet-b3654169.pth\n\n<span class=\"pl-c\"><span class=\"pl-c\">#</span> 8 GPU</span>\n./tools/dist_test.sh configs/glip/glip_atss_swin-t_fpn_dyhead_pretrain_obj365.py glip_tiny_a_mmdet-b3654169.pth 8</pre></div>\n<p>The result will be similar to this:</p>\n<div class=\"highlight highlight-source-shell\"><pre>Average Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>100 ] = 0.428\nAverage Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.50      <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.594\nAverage Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.75      <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.466\nAverage Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span> small <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.300\nAverage Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>medium <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.477\nAverage Precision  (AP) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span> large <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.534\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>100 ] = 0.634\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>300 ] = 0.634\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>   all <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.634\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span> small <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.473\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span>medium <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.690\nAverage Recall     (AR) @[ IoU<span class=\"pl-k\">=</span>0.50:0.95 <span class=\"pl-k\">|</span> area<span class=\"pl-k\">=</span> large <span class=\"pl-k\">|</span> maxDets<span class=\"pl-k\">=</span>1000 ] = 0.789</pre></div>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">XDecoder</h2><a id=\"user-content-xdecoder\" class=\"anchor\" aria-label=\"Permalink: XDecoder\" href=\"#xdecoder\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/cb126615-9402-4c19-8ea9-133722d7519c\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/cb126615-9402-4c19-8ea9-133722d7519c\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Installation</h2><a id=\"user-content-installation\" class=\"anchor\" aria-label=\"Permalink: Installation\" href=\"#installation\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c\"><span class=\"pl-c\">#</span> if source</span>\npip install -r requirements/multimodal.txt\n\n<span class=\"pl-c\"><span class=\"pl-c\">#</span> if wheel</span>\nmim install mmdet[multimodal]</pre></div>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">How to use it?</h2><a id=\"user-content-how-to-use-it\" class=\"anchor\" aria-label=\"Permalink: How to use it?\" href=\"#how-to-use-it\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>For convenience, you can download the weights to the <code>mmdetection</code> root dir</p>\n<div class=\"highlight highlight-source-shell\"><pre>wget https://download.openmmlab.com/mmdetection/v3.0/xdecoder/xdecoder_focalt_last_novg.pt\nwget https://download.openmmlab.com/mmdetection/v3.0/xdecoder/xdecoder_focalt_best_openseg.pt</pre></div>\n<p>The above two weights are directly copied from the official website without any modification. The specific source is <a href=\"https://github.com/microsoft/X-Decoder\">https://github.com/microsoft/X-Decoder</a></p>\n<p>For convenience of demonstration, please download <a href=\"https://github.com/microsoft/X-Decoder/tree/main/images\">the folder</a> and place it in the root directory of mmdetection.</p>\n<p><strong>(1) Open Vocabulary Semantic Segmentation</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/animals.png configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py --weights ../../xdecoder_focalt_last_novg.pt --texts zebra.giraffe</pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/c397c0ed-859a-4004-8725-78a591742bc8\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/c397c0ed-859a-4004-8725-78a591742bc8\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(2) Open Vocabulary Instance Segmentation</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/owls.jpeg configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py --weights ../../xdecoder_focalt_last_novg.pt --texts owl</pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/494b0b1c-4a42-4019-97ae-d33ee68af3d2\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/494b0b1c-4a42-4019-97ae-d33ee68af3d2\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(3) Open Vocabulary Panoptic Segmentation</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/street.jpg configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py --weights ../../xdecoder_focalt_last_novg.pt  --text car.person --stuff-text tree.sky</pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9ad1e0f4-75ce-4e37-a5cc-83e0e8a722ed\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/9ad1e0f4-75ce-4e37-a5cc-83e0e8a722ed\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(4) Referring Expression Segmentation</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/fruit.jpg configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py --weights ../../xdecoder_focalt_last_novg.pt  --text <span class=\"pl-s\"><span class=\"pl-pds\">\"</span>The larger watermelon. The front white flower. White tea pot.<span class=\"pl-pds\">\"</span></span></pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/f3ecdb50-20f0-4dc4-aa9c-90995ae04893\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/f3ecdb50-20f0-4dc4-aa9c-90995ae04893\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(5) Image Caption</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/penguin.jpeg configs/xdecoder-tiny_zeroshot_caption_coco2014.py --weights ../../xdecoder_focalt_last_novg.pt</pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/7690ab79-791e-4011-ab0c-01f46c4a3d80\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/7690ab79-791e-4011-ab0c-01f46c4a3d80\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(6) Referring Expression Image Caption</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/fruit.jpg configs/xdecoder-tiny_zeroshot_ref-caption.py --weights ../../xdecoder_focalt_last_novg.pt --text <span class=\"pl-s\"><span class=\"pl-pds\">'</span>White tea pot<span class=\"pl-pds\">'</span></span></pre></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/bae2fdba-0172-4fc8-8ad1-73b54c64ec30\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/bae2fdba-0172-4fc8-8ad1-73b54c64ec30\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p><strong>(7) Text Image Region Retrieval</strong></p>\n<div class=\"highlight highlight-source-shell\"><pre><span class=\"pl-c1\">cd</span> projects/XDecoder\npython demo.py ../../images/coco configs/xdecoder-tiny_zeroshot_text-image-retrieval.py --weights ../../xdecoder_focalt_last_novg.pt --text <span class=\"pl-s\"><span class=\"pl-pds\">'</span>pizza on the plate<span class=\"pl-pds\">'</span></span></pre></div>\n<pre lang=\"text\"><code>The image that best matches the given text is ../../images/coco/000.jpg and probability is 0.998\n</code></pre>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/479de6b2-88e7-41f0-8228-4b9a48f52954\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/479de6b2-88e7-41f0-8228-4b9a48f52954\" width=\"70%\" style=\"max-width: 100%;\"></a>\n</div>\n<p>We have also prepared a gradio program in the <code>projects/gradio_demo</code> directory, which you can run interactively all the inference supported by mmdetection in your browser.</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Models and results</h2><a id=\"user-content-models-and-results\" class=\"anchor\" aria-label=\"Permalink: Models and results\" href=\"#models-and-results\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Semantic segmentation on ADE20K</h3><a id=\"user-content-semantic-segmentation-on-ade20k\" class=\"anchor\" aria-label=\"Permalink: Semantic segmentation on ADE20K\" href=\"#semantic-segmentation-on-ade20k\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation\">docs</a>.</p>\n<p><strong>Test Command</strong></p>\n<p>Since semantic segmentation is a pixel-level task, we don't need to use a threshold to filter out low-confidence predictions. So we set <code>model.test_cfg.use_thr_for_mc=False</code> in the test command.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-semseg_ade20k.py xdecoder_focalt_best_openseg.pt 8 --cfg-options model.test_cfg.use_thr_for_mc=False</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">mIoU</th>\n<th align=\"center\">mIOU(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_best_openseg.pt</code></td>\n<td align=\"center\">25.24</td>\n<td align=\"center\">25.13</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-semseg_ade20k.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Instance segmentation on ADE20K</h3><a id=\"user-content-instance-segmentation-on-ade20k\" class=\"anchor\" aria-label=\"Permalink: Instance segmentation on ADE20K\" href=\"#instance-segmentation-on-ade20k\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation\">docs</a>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-instance_ade20k.py xdecoder_focalt_best_openseg.pt 8</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">mIoU</th>\n<th align=\"center\">mIOU(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_best_openseg.pt</code></td>\n<td align=\"center\">10.1</td>\n<td align=\"center\">10.1</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-instance_ade20k.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Panoptic segmentation on ADE20K</h3><a id=\"user-content-panoptic-segmentation-on-ade20k\" class=\"anchor\" aria-label=\"Permalink: Panoptic segmentation on ADE20K\" href=\"#panoptic-segmentation-on-ade20k\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#ade20k-2016-dataset-preparation\">docs</a>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_ade20k.py xdecoder_focalt_best_openseg.pt 8</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">mIoU</th>\n<th align=\"center\">mIOU(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_best_openseg.pt</code></td>\n<td align=\"center\">19.11</td>\n<td align=\"center\">18.97</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_ade20k.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Semantic segmentation on COCO2017</h3><a id=\"user-content-semantic-segmentation-on-coco2017\" class=\"anchor\" aria-label=\"Permalink: Semantic segmentation on COCO2017\" href=\"#semantic-segmentation-on-coco2017\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#coco-semantic-dataset-preparation\">docs</a> of <code>(2) use panoptic dataset</code> part.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py xdecoder_focalt_last_novg.pt 8 --cfg-options model.test_cfg.use_thr_for_mc=False</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">mIOU</th>\n<th align=\"center\">mIOU(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder-tiny_zeroshot_open-vocab-semseg_coco</code></td>\n<td align=\"center\">62.1</td>\n<td align=\"center\">62.1</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-semseg_coco.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Instance segmentation on COCO2017</h3><a id=\"user-content-instance-segmentation-on-coco2017\" class=\"anchor\" aria-label=\"Permalink: Instance segmentation on COCO2017\" href=\"#instance-segmentation-on-coco2017\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#basic-detection-dataset-preparation\">docs</a>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py xdecoder_focalt_last_novg.pt 8</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">Mask mAP</th>\n<th align=\"center\">Mask mAP(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder-tiny_zeroshot_open-vocab-instance_coco</code></td>\n<td align=\"center\">39.8</td>\n<td align=\"center\">39.7</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-instance_coco.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Panoptic segmentation on COCO2017</h3><a id=\"user-content-panoptic-segmentation-on-coco2017\" class=\"anchor\" aria-label=\"Permalink: Panoptic segmentation on COCO2017\" href=\"#panoptic-segmentation-on-coco2017\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#basic-detection-dataset-preparation\">docs</a>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py xdecoder_focalt_last_novg.pt 8</pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">PQ</th>\n<th align=\"center\">PQ(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder-tiny_zeroshot_open-vocab-panoptic_coco</code></td>\n<td align=\"center\">51.42</td>\n<td align=\"center\">51.16</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-panoptic_coco.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Referring segmentation on RefCOCO</h3><a id=\"user-content-referring-segmentation-on-refcoco\" class=\"anchor\" aria-label=\"Permalink: Referring segmentation on RefCOCO\" href=\"#referring-segmentation-on-refcoco\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#refcoco-dataset-preparation\">docs</a>.</p>\n<div class=\"highlight highlight-source-shell\"><pre>./tools/dist_test.sh  projects/XDecoder/configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py xdecoder_focalt_last_novg.pt 8  --cfg-options test_dataloader.dataset.split=<span class=\"pl-s\"><span class=\"pl-pds\">'</span>val<span class=\"pl-pds\">'</span></span></pre></div>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">text mode</th>\n<th align=\"center\">cIoU</th>\n<th align=\"center\">cIOU(official)</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_last_novg.pt</code></td>\n<td align=\"center\">select first</td>\n<td align=\"center\">58.8415</td>\n<td align=\"center\">57.85</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py\">config</a></td>\n</tr>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_last_novg.pt</code></td>\n<td align=\"center\">original</td>\n<td align=\"center\">60.0321</td>\n<td align=\"center\">-</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py\">config</a></td>\n</tr>\n<tr>\n<td align=\"left\"><code>xdecoder_focalt_last_novg.pt</code></td>\n<td align=\"center\">concat</td>\n<td align=\"center\">60.3551</td>\n<td align=\"center\">-</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_open-vocab-ref-seg_refcocog.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<p><strong>Note:</strong></p>\n<ol>\n<li>If you set the scale of <code>Resize</code> to (1024, 512), the result will be <code>57.69</code>.</li>\n<li>\n<code>text mode</code> is the <code>RefCoCoDataset</code> parameter in MMDetection, it determines the texts loaded to the data list. It can be set to <code>select_first</code>, <code>original</code>, <code>concat</code> and <code>random</code>.\n<ul>\n<li>\n<code>select_first</code>: select the first text in the text list as the description to an instance.</li>\n<li>\n<code>original</code>: use all texts in the text list as the description to an instance.</li>\n<li>\n<code>concat</code>: concatenate all texts in the text list as the description to an instance.</li>\n<li>\n<code>random</code>: randomly select one text in the text list as the description to an instance, usually used for training.</li>\n</ul>\n</li>\n</ol>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Image Caption on COCO2014</h3><a id=\"user-content-image-caption-on-coco2014\" class=\"anchor\" aria-label=\"Permalink: Image Caption on COCO2014\" href=\"#image-caption-on-coco2014\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>Prepare your dataset according to the <a href=\"../../docs/en/user_guides/dataset_prepare.md#coco-caption-dataset-preparation\">docs</a>.</p>\n<p>Before testing, you need to install jdk 1.8, otherwise it will prompt that java does not exist during the evaluation process</p>\n<pre><code>./tools/dist_test.sh projects/XDecoder/configs/xdecoder-tiny_zeroshot_caption_coco2014.py xdecoder_focalt_last_novg.pt 8\n</code></pre>\n<table>\n<thead>\n<tr>\n<th align=\"left\">Model</th>\n<th align=\"center\">BLEU-4</th>\n<th align=\"center\">CIDER</th>\n<th align=\"center\">Config</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td align=\"left\"><code>xdecoder-tiny_zeroshot_caption_coco2014</code></td>\n<td align=\"center\">35.26</td>\n<td align=\"center\">116.81</td>\n<td align=\"center\"><a href=\"configs/xdecoder-tiny_zeroshot_caption_coco2014.py\">config</a></td>\n</tr>\n</tbody>\n</table>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Gradio Demo</h2><a id=\"user-content-gradio-demo\" class=\"anchor\" aria-label=\"Permalink: Gradio Demo\" href=\"#gradio-demo\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<div align=\"center\">\n<a target=\"_blank\" rel=\"noopener noreferrer\" href=\"https://github.com/open-mmlab/mmdetection/assets/17425982/6c29886f-ae7a-4a55-8be4-352ee85b7d3e\"><img src=\"https://github.com/open-mmlab/mmdetection/assets/17425982/6c29886f-ae7a-4a55-8be4-352ee85b7d3e\" style=\"max-width: 100%;\"></a>\n</div>\n<p>Please refer to <a href=\"https://github.com/open-mmlab/mmdetection/blob/dev-3.x/projects/gradio_demo/README.md\">https://github.com/open-mmlab/mmdetection/blob/dev-3.x/projects/gradio_demo/README.md</a> for details.</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 30 developers contributed to this release.</p>\n<p>Thanks @jjjkkkjjj @lovelykite, @minato-ellie, @freepoet, @wufan-tb, @yalibian, @keyakiluo, @gihanjayatilaka, @i-aki-y, @xin-li-67, @RangeKing, @JingweiZhang12, @MambaWong, @lucianovk, @tall-josh, @xiuqhou, @jamiechoi1995, @YQisme, @yechenzhi, @bjzhb666, @xiexinch, @jamiechoi1995, @yarkable, @Renzhihan, @nijkah, @amaizr, @Lum1104, @zwhus, @Czm369, @hhaAndroid</p>\n","assets":[]},{"id":98310560,"tag":"v3.0.0","author":"hhaAndroid","name":"MMDetection releases v3.0.0","draft":false,"prerelease":false,"createdAt":"2023-04-06T06:55:28Z","publishedAt":"2023-04-06T07:26:03Z","markdown":"## v3.0.0 (6/4/2023)\r\n\r\nWe have released the official version of MMDetection v3.0.0\r\n\r\n### Highlights\r\n\r\n- Support Semi-automatic annotation Base [Label-Studio](https://github.com/open-mmlab/mmdetection/blob/main/projects/LabelStudio) (#10039)\r\n- Support [EfficientDet](https://github.com/open-mmlab/mmdetection/blob/main/projects/EfficientDet) in projects (#9810)\r\n\r\n### New Features\r\n\r\n- File I/O migration and reconstruction (#9709)\r\n- Release DINO Swin-L 36e model (#9927)\r\n\r\n### Bug Fixes\r\n\r\n- Fix benchmark script (#9865)\r\n- Fix the crop method of PolygonMasks (#9858)\r\n- Fix Albu augmentation with the mask shape (#9918)\r\n- Fix `RTMDetIns` prior generator device error (#9964)\r\n- Fix `img_shape` in data pipeline (#9966)\r\n- Fix cityscapes import error (#9984)\r\n- Fix `solov2_r50_fpn_ms-3x_coco.py` config error (#10030)\r\n- Fix Conditional DETR AP and Log (#9889)\r\n- Fix accepting an unexpected argument local-rank in PyTorch 2.0 (#10050)\r\n- Fix `common/ms_3x_coco-instance.py` config error (#10056)\r\n- Fix compute flops error (#10051)\r\n- Delete `data_root` in `CocoOccludedSeparatedMetric` to fix bug (#9969)\r\n- Unifying metafile.yml (#9849)\r\n\r\n### Improvements\r\n\r\n- Added BoxInst r101 config (#9967)\r\n- Added config migration guide (#9960)\r\n- Added more social networking links (#10021)\r\n- Added RTMDet config introduce (#10042)\r\n- Added visualization docs (#9938, #10058)\r\n- Refined data_prepare docs (#9935)\r\n- Added support for setting the cache_size_limit parameter of dynamo in PyTorch 2.0 (#10054)\r\n- Updated coco_metric.py (#10033)\r\n- Update type hint (#10040)\r\n\r\n### Contributors\r\n\r\nA total of 19 developers contributed to this release.\r\n\r\nThanks @IRONICBo, @vansin, @RangeKing, @Ghlerrix, @okotaku, @JosonChan1998, @zgzhengSEU, @bobo0810, @yechenzhi, @Zheng-LinXiao, @LYMDLUT, @yarkable, @xiejiajiannb, @chhluo, @BIGWangYuDong, @RangiLyu, @zwhus, @hhaAndroid, @ZwwWayne\r\n","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">v3.0.0 (6/4/2023)</h2><a id=\"user-content-v300-642023\" class=\"anchor\" aria-label=\"Permalink: v3.0.0 (6/4/2023)\" href=\"#v300-642023\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>We have released the official version of MMDetection v3.0.0</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support Semi-automatic annotation Base <a href=\"https://github.com/open-mmlab/mmdetection/blob/main/projects/LabelStudio\">Label-Studio</a> (#10039)</li>\n<li>Support <a href=\"https://github.com/open-mmlab/mmdetection/blob/main/projects/EfficientDet\">EfficientDet</a> in projects (#9810)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>File I/O migration and reconstruction (#9709)</li>\n<li>Release DINO Swin-L 36e model (#9927)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix benchmark script (#9865)</li>\n<li>Fix the crop method of PolygonMasks (#9858)</li>\n<li>Fix Albu augmentation with the mask shape (#9918)</li>\n<li>Fix <code>RTMDetIns</code> prior generator device error (#9964)</li>\n<li>Fix <code>img_shape</code> in data pipeline (#9966)</li>\n<li>Fix cityscapes import error (#9984)</li>\n<li>Fix <code>solov2_r50_fpn_ms-3x_coco.py</code> config error (#10030)</li>\n<li>Fix Conditional DETR AP and Log (#9889)</li>\n<li>Fix accepting an unexpected argument local-rank in PyTorch 2.0 (#10050)</li>\n<li>Fix <code>common/ms_3x_coco-instance.py</code> config error (#10056)</li>\n<li>Fix compute flops error (#10051)</li>\n<li>Delete <code>data_root</code> in <code>CocoOccludedSeparatedMetric</code> to fix bug (#9969)</li>\n<li>Unifying metafile.yml (#9849)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Added BoxInst r101 config (#9967)</li>\n<li>Added config migration guide (#9960)</li>\n<li>Added more social networking links (#10021)</li>\n<li>Added RTMDet config introduce (#10042)</li>\n<li>Added visualization docs (#9938, #10058)</li>\n<li>Refined data_prepare docs (#9935)</li>\n<li>Added support for setting the cache_size_limit parameter of dynamo in PyTorch 2.0 (#10054)</li>\n<li>Updated coco_metric.py (#10033)</li>\n<li>Update type hint (#10040)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 19 developers contributed to this release.</p>\n<p>Thanks @IRONICBo, @vansin, @RangeKing, @Ghlerrix, @okotaku, @JosonChan1998, @zgzhengSEU, @bobo0810, @yechenzhi, @Zheng-LinXiao, @LYMDLUT, @yarkable, @xiejiajiannb, @chhluo, @BIGWangYuDong, @RangiLyu, @zwhus, @hhaAndroid, @ZwwWayne</p>\n","assets":[]},{"id":93875923,"tag":"v2.28.2","author":"ZwwWayne","name":"MMDetection V2.28.2 Release","draft":false,"prerelease":false,"createdAt":"2023-02-28T02:17:10Z","publishedAt":"2023-02-28T02:24:52Z","markdown":"## New Features and Improvements\r\n\r\n- Add Twitter, Discord, Medium and YouTube link (#9774)\r\n- Update `customize_runtime.md` (#9797)\r\n\r\n## Bug Fixes\r\n\r\n- Fix `WIDERFace SSD` loss for Nan problem (#9734)\r\n- Fix missing API documentation in Readthedoc (#9729)\r\n- Fix the configuration file and log path of CenterNet (#9791)\r\n\r\n\r\n## New Contributors\r\n* @Ginray made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9791\r\n* @co63oc made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9797\r\n\r\n## Contributors\r\n\r\nA total of 4 developers contributed to this release.\r\nThanks @co63oc, @Ginray, @vansin, @RangiLyu\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.28.1...v2.28.2","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features and Improvements</h2><a id=\"user-content-new-features-and-improvements\" class=\"anchor\" aria-label=\"Permalink: New Features and Improvements\" href=\"#new-features-and-improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Add Twitter, Discord, Medium and YouTube link (#9774)</li>\n<li>Update <code>customize_runtime.md</code> (#9797)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix <code>WIDERFace SSD</code> loss for Nan problem (#9734)</li>\n<li>Fix missing API documentation in Readthedoc (#9729)</li>\n<li>Fix the configuration file and log path of CenterNet (#9791)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@Ginray made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9791\">https://github.com/open-mmlab/mmdetection/pull/9791</a>\n</li>\n<li>@co63oc made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9797\">https://github.com/open-mmlab/mmdetection/pull/9797</a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 4 developers contributed to this release.\nThanks @co63oc, @Ginray, @vansin, @RangiLyu</p>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.28.1...v2.28.2\">https://github.com/open-mmlab/mmdetection/compare/v2.28.1...v2.28.2</a></p>\n","assets":[]},{"id":93812953,"tag":"v3.0.0rc6","author":"ZwwWayne","name":"MMDetection V3.0.0rc6 Release","draft":false,"prerelease":true,"createdAt":"2023-02-27T15:39:56Z","publishedAt":"2023-02-27T16:00:04Z","markdown":"## Highlights\r\n\r\n- Support [Boxinst](../../../configs/boxinst), [Objects365 Dataset](../../../configs/objects365), and [Separated and Occluded COCO metric](../user_guides/useful_tools.md#COCO-Separated-&-Occluded-Mask-Metric)\r\n- Support [ConvNeXt-V2](../../../projects/ConvNeXt-V2), [DiffusionDet](../../../projects/DiffusionDet), and inference of [EfficientDet](../../../projects/EfficientDet) and [Detic](../../../projects/Detic) in `Projects`\r\n- Refactor [DETR](../../../configs/detr) series and support [Conditional-DETR](../../../configs/conditional_detr), [DAB-DETR](../../../configs/dab_detr), and [DINO](../../../configs/detr)\r\n- Support `DetInferencer` for inference, Test Time Augmentation, and automatically importing modules from registry\r\n- Support RTMDet-Ins ONNXRuntime and TensorRT [deployment](../../../configs/rtmdet/README.md#deployment-tutorial)\r\n- Support [calculating FLOPs of detectors](../user_guides/useful_tools.md#Model-Complexity)\r\n\r\n## New Features\r\n\r\n- Support [Boxinst](https://arxiv.org/abs/2012.02310) (#9525)\r\n- Support [Objects365 Dataset](https://openaccess.thecvf.com/content_ICCV_2019/papers/Shao_Objects365_A_Large-Scale_High-Quality_Dataset_for_Object_Detection_ICCV_2019_paper.pdf) (#9600)\r\n- Support [ConvNeXt-V2](http://arxiv.org/abs/2301.00808) in `Projects` (#9619)\r\n- Support [DiffusionDet](https://arxiv.org/abs/2211.09788) in `Projects` (#9639, #9768)\r\n- Support [Detic](http://arxiv.org/abs/2201.02605) inference in `Projects` (#9645)\r\n- Support [EfficientDet](https://arxiv.org/abs/1911.09070) inference in `Projects` (#9645)\r\n- Support [Separated and Occluded COCO metric](https://arxiv.org/abs/2210.10046) (#9710)\r\n- Support auto import modules from registry (#9143)\r\n- Refactor DETR series and support Conditional-DETR, DAB-DETR and DINO (#9646)\r\n- Support `DetInferencer` for inference (#9561)\r\n- Support Test Time Augmentation (#9452)\r\n- Support calculating FLOPs of detectors (#9777)\r\n\r\n## Bug Fixes\r\n\r\n- Fix deprecating old type alias due to new version of numpy (#9625, #9537)\r\n- Fix VOC metrics (#9784)\r\n- Fix the wrong link of RTMDet-x log (#9549)\r\n- Fix RTMDet link in README (#9575)\r\n- Fix MMDet get flops error (#9589)\r\n- Fix `use_depthwise` in RTMDet (#9624)\r\n- Fix `albumentations` augmentation post process with masks (#9551)\r\n- Fix DETR series Unit Test (#9647)\r\n- Fix `LoadPanopticAnnotations` bug (#9703)\r\n- Fix `isort` CI (#9680)\r\n- Fix amp pooling overflow (#9670)\r\n- Fix docstring about noise in DINO (#9747)\r\n- Fix potential bug in `MultiImageMixDataset` (#9764)\r\n\r\n## Improvements\r\n\r\n- Replace NumPy transpose with PyTorch permute to speed-up (#9762)\r\n- Deprecate `sklearn` (#9725)\r\n- Add RTMDet-Ins deployment guide (#9823)\r\n- Update RTMDet config and README (#9603)\r\n- Replace the models used in the tutorial document with RTMDet (#9843)\r\n- Adjust the minimum supported python version to 3.7 (#9602)\r\n- Support modifying palette through configuration (#9445)\r\n- Update README document in `Project` (#9599)\r\n- Replace `github` with `gitee` in `.pre-commit-config-zh-cn.yaml` file (#9586)\r\n- Use official `isort` in `.pre-commit-config.yaml` file (#9701)\r\n- Change MMCV minimum version to `2.0.0rc4` for `dev-3.x` (#9695)\r\n- Add Chinese version of single_stage_as_rpn.md and test_results_submission.md (#9434)\r\n- Add OpenDataLab download link (#9605, #9738)\r\n- Add type hints of several layers (#9346)\r\n- Add typehint for `DarknetBottleneck` (#9591)\r\n- Add dockerfile that is easier to use in China (#9659)\r\n- Add twitter, discord, medium, and youtube link (#9775)\r\n- Prepare for merging refactor-detr (#9656)\r\n- Add metafile to ConditionalDETR, DABDETR and DINO (#9715)\r\n- Support to modify `non_blocking` parameters (#9723)\r\n- Comment repeater visualizer register (#9740)\r\n- Update user guide: `finetune.md` and `inference.md` (#9578)\r\n\r\n## New Contributors\r\n\r\n- @NoFish-528 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9346>\r\n- @137208 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9434>\r\n- @lyviva made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9625>\r\n- @zwhus made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9589>\r\n- @zylo117 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9670>\r\n- @chg0901 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9740>\r\n- @DanShouzhu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9578\r\n\r\n## Contributors\r\n\r\nA total of 27 developers contributed to this release.\r\n\r\nThanks @JosonChan1998, @RangeKing, @NoFish-528, @likyoo, @Xiangxu-0103, @137208, @PeterH0323, @tianleiSHI, @wufan-tb, @lyviva, @zwhus, @jshilong, @Li-Qingyun, @sanbuphy, @zylo117, @triple-Mu, @KeiChiTse, @LYMDLUT, @nijkah, @chg0901, @DanShouzhu, @zytx121, @vansin, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne\r\n\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc5...v3.0.0rc6","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"../../../configs/boxinst\">Boxinst</a>, <a href=\"../../../configs/objects365\">Objects365 Dataset</a>, and <a href=\"../user_guides/useful_tools.md#COCO-Separated-&amp;-Occluded-Mask-Metric\">Separated and Occluded COCO metric</a>\n</li>\n<li>Support <a href=\"../../../projects/ConvNeXt-V2\">ConvNeXt-V2</a>, <a href=\"../../../projects/DiffusionDet\">DiffusionDet</a>, and inference of <a href=\"../../../projects/EfficientDet\">EfficientDet</a> and <a href=\"../../../projects/Detic\">Detic</a> in <code>Projects</code>\n</li>\n<li>Refactor <a href=\"../../../configs/detr\">DETR</a> series and support <a href=\"../../../configs/conditional_detr\">Conditional-DETR</a>, <a href=\"../../../configs/dab_detr\">DAB-DETR</a>, and <a href=\"../../../configs/detr\">DINO</a>\n</li>\n<li>Support <code>DetInferencer</code> for inference, Test Time Augmentation, and automatically importing modules from registry</li>\n<li>Support RTMDet-Ins ONNXRuntime and TensorRT <a href=\"../../../configs/rtmdet/README.md#deployment-tutorial\">deployment</a>\n</li>\n<li>Support <a href=\"../user_guides/useful_tools.md#Model-Complexity\">calculating FLOPs of detectors</a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2012.02310\" rel=\"nofollow\">Boxinst</a> (#9525)</li>\n<li>Support <a href=\"https://openaccess.thecvf.com/content_ICCV_2019/papers/Shao_Objects365_A_Large-Scale_High-Quality_Dataset_for_Object_Detection_ICCV_2019_paper.pdf\" rel=\"nofollow\">Objects365 Dataset</a> (#9600)</li>\n<li>Support <a href=\"http://arxiv.org/abs/2301.00808\" rel=\"nofollow\">ConvNeXt-V2</a> in <code>Projects</code> (#9619)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2211.09788\" rel=\"nofollow\">DiffusionDet</a> in <code>Projects</code> (#9639, #9768)</li>\n<li>Support <a href=\"http://arxiv.org/abs/2201.02605\" rel=\"nofollow\">Detic</a> inference in <code>Projects</code> (#9645)</li>\n<li>Support <a href=\"https://arxiv.org/abs/1911.09070\" rel=\"nofollow\">EfficientDet</a> inference in <code>Projects</code> (#9645)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2210.10046\" rel=\"nofollow\">Separated and Occluded COCO metric</a> (#9710)</li>\n<li>Support auto import modules from registry (#9143)</li>\n<li>Refactor DETR series and support Conditional-DETR, DAB-DETR and DINO (#9646)</li>\n<li>Support <code>DetInferencer</code> for inference (#9561)</li>\n<li>Support Test Time Augmentation (#9452)</li>\n<li>Support calculating FLOPs of detectors (#9777)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix deprecating old type alias due to new version of numpy (#9625, #9537)</li>\n<li>Fix VOC metrics (#9784)</li>\n<li>Fix the wrong link of RTMDet-x log (#9549)</li>\n<li>Fix RTMDet link in README (#9575)</li>\n<li>Fix MMDet get flops error (#9589)</li>\n<li>Fix <code>use_depthwise</code> in RTMDet (#9624)</li>\n<li>Fix <code>albumentations</code> augmentation post process with masks (#9551)</li>\n<li>Fix DETR series Unit Test (#9647)</li>\n<li>Fix <code>LoadPanopticAnnotations</code> bug (#9703)</li>\n<li>Fix <code>isort</code> CI (#9680)</li>\n<li>Fix amp pooling overflow (#9670)</li>\n<li>Fix docstring about noise in DINO (#9747)</li>\n<li>Fix potential bug in <code>MultiImageMixDataset</code> (#9764)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Replace NumPy transpose with PyTorch permute to speed-up (#9762)</li>\n<li>Deprecate <code>sklearn</code> (#9725)</li>\n<li>Add RTMDet-Ins deployment guide (#9823)</li>\n<li>Update RTMDet config and README (#9603)</li>\n<li>Replace the models used in the tutorial document with RTMDet (#9843)</li>\n<li>Adjust the minimum supported python version to 3.7 (#9602)</li>\n<li>Support modifying palette through configuration (#9445)</li>\n<li>Update README document in <code>Project</code> (#9599)</li>\n<li>Replace <code>github</code> with <code>gitee</code> in <code>.pre-commit-config-zh-cn.yaml</code> file (#9586)</li>\n<li>Use official <code>isort</code> in <code>.pre-commit-config.yaml</code> file (#9701)</li>\n<li>Change MMCV minimum version to <code>2.0.0rc4</code> for <code>dev-3.x</code> (#9695)</li>\n<li>Add Chinese version of single_stage_as_rpn.md and test_results_submission.md (#9434)</li>\n<li>Add OpenDataLab download link (#9605, #9738)</li>\n<li>Add type hints of several layers (#9346)</li>\n<li>Add typehint for <code>DarknetBottleneck</code> (#9591)</li>\n<li>Add dockerfile that is easier to use in China (#9659)</li>\n<li>Add twitter, discord, medium, and youtube link (#9775)</li>\n<li>Prepare for merging refactor-detr (#9656)</li>\n<li>Add metafile to ConditionalDETR, DABDETR and DINO (#9715)</li>\n<li>Support to modify <code>non_blocking</code> parameters (#9723)</li>\n<li>Comment repeater visualizer register (#9740)</li>\n<li>Update user guide: <code>finetune.md</code> and <code>inference.md</code> (#9578)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@NoFish-528 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9346\">https://github.com/open-mmlab/mmdetection/pull/9346</a>\n</li>\n<li>@137208 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9434\">https://github.com/open-mmlab/mmdetection/pull/9434</a>\n</li>\n<li>@lyviva made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9625\">https://github.com/open-mmlab/mmdetection/pull/9625</a>\n</li>\n<li>@zwhus made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9589\">https://github.com/open-mmlab/mmdetection/pull/9589</a>\n</li>\n<li>@zylo117 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9670\">https://github.com/open-mmlab/mmdetection/pull/9670</a>\n</li>\n<li>@chg0901 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9740\">https://github.com/open-mmlab/mmdetection/pull/9740</a>\n</li>\n<li>@DanShouzhu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9578\">https://github.com/open-mmlab/mmdetection/pull/9578</a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 27 developers contributed to this release.</p>\n<p>Thanks @JosonChan1998, @RangeKing, @NoFish-528, @likyoo, @Xiangxu-0103, @137208, @PeterH0323, @tianleiSHI, @wufan-tb, @lyviva, @zwhus, @jshilong, @Li-Qingyun, @sanbuphy, @zylo117, @triple-Mu, @KeiChiTse, @LYMDLUT, @nijkah, @chg0901, @DanShouzhu, @zytx121, @vansin, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne</p>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc5...v3.0.0rc6\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc5...v3.0.0rc6</a></p>\n","assets":[]},{"id":90964401,"tag":"v2.28.1","author":"RangiLyu","name":"MMDetection V2.28.1 Release","draft":false,"prerelease":false,"createdAt":"2023-02-01T11:29:57Z","publishedAt":"2023-02-01T12:18:43Z","markdown":"### Bug Fixes\r\n\r\n- Enable to set float mlp_ratio in SwinTransformer (#8670)\r\n- Fix import error that causes training failure (#9694)\r\n- Fix isort version in lint (#9685)\r\n- Fix init_cfg of YOLOF (#8243)\r\n\r\n### Contributors\r\n\r\nA total of 4 developers contributed to this release.\r\nThanks @triple-Mu, @i-aki-y, @twmht, @RangiLyu\r\n\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.28.0...v2.28.1","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Enable to set float mlp_ratio in SwinTransformer (#8670)</li>\n<li>Fix import error that causes training failure (#9694)</li>\n<li>Fix isort version in lint (#9685)</li>\n<li>Fix init_cfg of YOLOF (#8243)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 4 developers contributed to this release.\nThanks @triple-Mu, @i-aki-y, @twmht, @RangiLyu</p>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.28.0...v2.28.1\">https://github.com/open-mmlab/mmdetection/compare/v2.28.0...v2.28.1</a></p>\n","assets":[]},{"id":90521253,"tag":"v2.28.0","author":"ZwwWayne","name":"MMDetection V2.28.0 Release","draft":false,"prerelease":false,"createdAt":"2023-01-28T11:57:08Z","publishedAt":"2023-01-28T12:58:14Z","markdown":"### Highlights\r\n\r\n- Support Objects365 Dataset and Separated and Occluded COCO metric\r\n- Support acceleration of RetinaNet and SSD on Ascend\r\n- Deprecate the support of Python 3.6\r\n\r\n### New Features and Improvements\r\n\r\n- Support Objects365 Dataset (#7525)\r\n- Support [Separated and Occluded COCO metric](https://arxiv.org/abs/2210.10046) (#9574)\r\n- Support acceleration of RetinaNet and SSD on Ascend with documentation (#9648, #9614)\r\n- Added missing `-` to `--format-only` in documentation.\r\n\r\n### Deprecations\r\n\r\n- Upgrade the minimum Python version to 3.7, the support of Python 3.6 is no longer guaranteed (#9604)\r\n\r\n### Bug Fixes\r\n\r\n- Fix validation loss logging by (#9663)\r\n- Fix inconsistent float precision between mmdet and mmcv (#9570)\r\n- Fix argument name for fp32 in `DeformableDETRHead` (#9607)\r\n- Fix typo of all config file path in Metafile.yml (#9627)\r\n\r\n### Contributors\r\n\r\nA total of 11 developers contributed to this release.\r\nThanks @eantono, @akstt, @@lpizzinidev, @RangiLyu, @kbumsik, @tianleiSHI, @nijkah, @BIGWangYuDong, @wangjiangben-hw, @@jamiechoi1995, @ZwwWayne\r\n\r\n## New Contributors\r\n* @kbumsik made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9627\r\n* @akstt made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9614\r\n* @lpizzinidev made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9649\r\n* @eantono made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9663\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.27.0...v2.28.0","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support Objects365 Dataset and Separated and Occluded COCO metric</li>\n<li>Support acceleration of RetinaNet and SSD on Ascend</li>\n<li>Deprecate the support of Python 3.6</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features and Improvements</h3><a id=\"user-content-new-features-and-improvements\" class=\"anchor\" aria-label=\"Permalink: New Features and Improvements\" href=\"#new-features-and-improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support Objects365 Dataset (#7525)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2210.10046\" rel=\"nofollow\">Separated and Occluded COCO metric</a> (#9574)</li>\n<li>Support acceleration of RetinaNet and SSD on Ascend with documentation (#9648, #9614)</li>\n<li>Added missing <code>-</code> to <code>--format-only</code> in documentation.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Deprecations</h3><a id=\"user-content-deprecations\" class=\"anchor\" aria-label=\"Permalink: Deprecations\" href=\"#deprecations\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Upgrade the minimum Python version to 3.7, the support of Python 3.6 is no longer guaranteed (#9604)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix validation loss logging by (#9663)</li>\n<li>Fix inconsistent float precision between mmdet and mmcv (#9570)</li>\n<li>Fix argument name for fp32 in <code>DeformableDETRHead</code> (#9607)</li>\n<li>Fix typo of all config file path in Metafile.yml (#9627)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 11 developers contributed to this release.\nThanks @eantono, @akstt, @@lpizzinidev, @RangiLyu, @kbumsik, @tianleiSHI, @nijkah, @BIGWangYuDong, @wangjiangben-hw, @@jamiechoi1995, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@kbumsik made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9627\">https://github.com/open-mmlab/mmdetection/pull/9627</a>\n</li>\n<li>@akstt made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9614\">https://github.com/open-mmlab/mmdetection/pull/9614</a>\n</li>\n<li>@lpizzinidev made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9649\">https://github.com/open-mmlab/mmdetection/pull/9649</a>\n</li>\n<li>@eantono made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9663\">https://github.com/open-mmlab/mmdetection/pull/9663</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.27.0...v2.28.0\">https://github.com/open-mmlab/mmdetection/compare/v2.27.0...v2.28.0</a></p>\n","assets":[]},{"id":87975504,"tag":"v2.27.0","author":"RangiLyu","name":"MMDetection V2.27.0 Release","draft":false,"prerelease":false,"createdAt":"2023-01-05T07:27:03Z","publishedAt":"2023-01-05T10:26:05Z","markdown":"## Highlights\r\n\r\n- Support receptive field search of CNN models([TPAMI 2022: RF-Next](http://mftp.mmcheng.net/Papers/22TPAMI-ActionSeg.pdf)) (#8191)\r\n\r\n## Bug Fixes\r\n\r\n- Fix deadlock issue related with MMDetWandbHook (#9476)\r\n\r\n## Improvements\r\n\r\n- Add minimum GitHub token permissions for workflows (#8928)\r\n- Delete compatible code for parrots in roi extractor (#9503)\r\n- Deprecate np.bool Type Alias (#9498)\r\n- Replace numpy transpose with torch permute to speed-up data pre-processing (#9533)\r\n\r\n## Documents\r\n\r\n- Fix typo in docs/zh_cn/tutorials/config.md (#9416)\r\n- Fix Faster RCNN FP16 config link in README (#9366)\r\n\r\n## Contributors\r\n\r\nA total of 12 developers contributed to this release.\r\nThanks @Min-Sheng, @gasvn, @lzyhha, @jbwang1997, @zachcoleman, @chenyuwang814, @MilkClouds, @Fizzez, @boahc077, @apatsekin, @zytx121, @DonggeunYu\r\n\r\n## New Contributors\r\n* @DonggeunYu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9416\r\n* @apatsekin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9366\r\n* @boahc077 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8928\r\n* @MilkClouds made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9476\r\n* @chenyuwang814 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9503\r\n* @zachcoleman made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9498\r\n* @Min-Sheng made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9533\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.26.0...v2.27.0","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support receptive field search of CNN models(<a href=\"http://mftp.mmcheng.net/Papers/22TPAMI-ActionSeg.pdf\" rel=\"nofollow\">TPAMI 2022: RF-Next</a>) (#8191)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix deadlock issue related with MMDetWandbHook (#9476)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Add minimum GitHub token permissions for workflows (#8928)</li>\n<li>Delete compatible code for parrots in roi extractor (#9503)</li>\n<li>Deprecate np.bool Type Alias (#9498)</li>\n<li>Replace numpy transpose with torch permute to speed-up data pre-processing (#9533)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Documents</h2><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix typo in docs/zh_cn/tutorials/config.md (#9416)</li>\n<li>Fix Faster RCNN FP16 config link in README (#9366)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 12 developers contributed to this release.\nThanks @Min-Sheng, @gasvn, @lzyhha, @jbwang1997, @zachcoleman, @chenyuwang814, @MilkClouds, @Fizzez, @boahc077, @apatsekin, @zytx121, @DonggeunYu</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@DonggeunYu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9416\">https://github.com/open-mmlab/mmdetection/pull/9416</a>\n</li>\n<li>@apatsekin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9366\">https://github.com/open-mmlab/mmdetection/pull/9366</a>\n</li>\n<li>@boahc077 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8928\">https://github.com/open-mmlab/mmdetection/pull/8928</a>\n</li>\n<li>@MilkClouds made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9476\">https://github.com/open-mmlab/mmdetection/pull/9476</a>\n</li>\n<li>@chenyuwang814 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9503\">https://github.com/open-mmlab/mmdetection/pull/9503</a>\n</li>\n<li>@zachcoleman made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9498\">https://github.com/open-mmlab/mmdetection/pull/9498</a>\n</li>\n<li>@Min-Sheng made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9533\">https://github.com/open-mmlab/mmdetection/pull/9533</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.26.0...v2.27.0\">https://github.com/open-mmlab/mmdetection/compare/v2.26.0...v2.27.0</a></p>\n","assets":[]},{"id":87302916,"tag":"v3.0.0rc5","author":"ZwwWayne","name":"MMDetection V3.0.0rc5 Release","draft":false,"prerelease":true,"createdAt":"2022-12-27T13:13:14Z","publishedAt":"2022-12-27T13:25:21Z","markdown":"## Highlights\r\n\r\n- Support [RTMDet](https://arxiv.org/abs/2212.07784) instance segmentation models. The technical report of RTMDet is on [arxiv](https://arxiv.org/abs/2212.07784)\r\n- Support SSHContextModule in paper [SSH: Single Stage Headless Face Detector](https://arxiv.org/abs/1708.03979).\r\n\r\n## New Features\r\n\r\n- Support [RTMDet](https://arxiv.org/abs/2212.07784) instance segmentation models and improve RTMDet test config (#9494)\r\n- Support SSHContextModule in paper [SSH: Single Stage Headless Face Detector](https://arxiv.org/abs/1708.03979) (#8953)\r\n- Release [CondInst](https://arxiv.org/abs/2003.05664) pre-trained model (#9406)\r\n\r\n## Bug Fixes\r\n\r\n- Fix CondInst predict error when `batch_size` is greater than 1 in inference (#9400)\r\n- Fix the bug of visualization when the dtype of the pipeline output image is not uint8 in browse dataset (#9401)\r\n- Fix `analyze_logs.py` to plot mAP and calculate train time correctly (#9409)\r\n- Fix backward inplace error with `PAFPN` (#9450)\r\n- Fix config import links in model converters (#9441)\r\n- Fix `DeformableDETRHead` object has no attribute `loss_single` (#9477)\r\n- Fix the logic of pseudo bboxes predicted by teacher model in SemiBaseDetector (#9414)\r\n- Fix demo API in instance segmentation tutorial (#9226)\r\n- Fix `analyze_results` (#9380)\r\n- Fix the error that Readthedocs API cannot be displayed (#9510)\r\n\r\n## Improvements\r\n\r\n- Remove legacy `builder.py` (#9479)\r\n- Make sure the pipeline argument shape is in `(width, height)` order (#9324)\r\n- Add `.pre-commit-config-zh-cn.yaml` file (#9388)\r\n- Refactor dataset metainfo to lowercase (#9469)\r\n- Add PyTorch 1.13 checking in CI (#9478)\r\n- Adjust `FocalLoss` and `QualityFocalLoss` to allow different kinds of targets (#9481)\r\n- Refactor `setup.cfg` (#9370)\r\n- Clip saturation value to valid range `[0, 1]` (#9391)\r\n- Only keep meta and state_dict when publishing model (#9356)\r\n- Add segm evaluator in ms-poly_3x_coco_instance config (#9524)\r\n- Update deployment guide (#9527)\r\n- Update zh_cn `faq.md` (#9396)\r\n- Update `get_started` (#9480)\r\n- Update the zh_cn user_guides of `useful_tools.md` and `useful_hooks.md` (#9453)\r\n- Add type hints for `bfp` and `channel_mapper` (#9410)\r\n- Add type hints of several losses (#9397)\r\n- Add type hints and update docstring for task modules (#9468)\r\n\r\n## Contributors\r\n\r\nA total of 20 developers contributed to this release.\r\n\r\nThanks @liuyanyi, @RangeKing, @lihua199710, @MambaWong, @sanbuphy, @Xiangxu-0103, @twmht, @JunyaoHu, @Chan-Sun, @tianleiSHI, @zytx121, @kitecats, @QJC123654, @JosonChan1998, @lvhan028, @Czm369, @BIGWangYuDong, @RangiLyu, @hhaAndroid, @ZwwWayne\r\n\r\n## New Contributors\r\n* @lihua199710 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9388\r\n* @twmht made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9450\r\n* @tianleiSHI made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9453\r\n* @kitecats made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9481\r\n* @QJC123654 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9468\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc4...v3.0.0rc5","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2212.07784\" rel=\"nofollow\">RTMDet</a> instance segmentation models. The technical report of RTMDet is on <a href=\"https://arxiv.org/abs/2212.07784\" rel=\"nofollow\">arxiv</a>\n</li>\n<li>Support SSHContextModule in paper <a href=\"https://arxiv.org/abs/1708.03979\" rel=\"nofollow\">SSH: Single Stage Headless Face Detector</a>.</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2212.07784\" rel=\"nofollow\">RTMDet</a> instance segmentation models and improve RTMDet test config (#9494)</li>\n<li>Support SSHContextModule in paper <a href=\"https://arxiv.org/abs/1708.03979\" rel=\"nofollow\">SSH: Single Stage Headless Face Detector</a> (#8953)</li>\n<li>Release <a href=\"https://arxiv.org/abs/2003.05664\" rel=\"nofollow\">CondInst</a> pre-trained model (#9406)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix CondInst predict error when <code>batch_size</code> is greater than 1 in inference (#9400)</li>\n<li>Fix the bug of visualization when the dtype of the pipeline output image is not uint8 in browse dataset (#9401)</li>\n<li>Fix <code>analyze_logs.py</code> to plot mAP and calculate train time correctly (#9409)</li>\n<li>Fix backward inplace error with <code>PAFPN</code> (#9450)</li>\n<li>Fix config import links in model converters (#9441)</li>\n<li>Fix <code>DeformableDETRHead</code> object has no attribute <code>loss_single</code> (#9477)</li>\n<li>Fix the logic of pseudo bboxes predicted by teacher model in SemiBaseDetector (#9414)</li>\n<li>Fix demo API in instance segmentation tutorial (#9226)</li>\n<li>Fix <code>analyze_results</code> (#9380)</li>\n<li>Fix the error that Readthedocs API cannot be displayed (#9510)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Remove legacy <code>builder.py</code> (#9479)</li>\n<li>Make sure the pipeline argument shape is in <code>(width, height)</code> order (#9324)</li>\n<li>Add <code>.pre-commit-config-zh-cn.yaml</code> file (#9388)</li>\n<li>Refactor dataset metainfo to lowercase (#9469)</li>\n<li>Add PyTorch 1.13 checking in CI (#9478)</li>\n<li>Adjust <code>FocalLoss</code> and <code>QualityFocalLoss</code> to allow different kinds of targets (#9481)</li>\n<li>Refactor <code>setup.cfg</code> (#9370)</li>\n<li>Clip saturation value to valid range <code>[0, 1]</code> (#9391)</li>\n<li>Only keep meta and state_dict when publishing model (#9356)</li>\n<li>Add segm evaluator in ms-poly_3x_coco_instance config (#9524)</li>\n<li>Update deployment guide (#9527)</li>\n<li>Update zh_cn <code>faq.md</code> (#9396)</li>\n<li>Update <code>get_started</code> (#9480)</li>\n<li>Update the zh_cn user_guides of <code>useful_tools.md</code> and <code>useful_hooks.md</code> (#9453)</li>\n<li>Add type hints for <code>bfp</code> and <code>channel_mapper</code> (#9410)</li>\n<li>Add type hints of several losses (#9397)</li>\n<li>Add type hints and update docstring for task modules (#9468)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 20 developers contributed to this release.</p>\n<p>Thanks @liuyanyi, @RangeKing, @lihua199710, @MambaWong, @sanbuphy, @Xiangxu-0103, @twmht, @JunyaoHu, @Chan-Sun, @tianleiSHI, @zytx121, @kitecats, @QJC123654, @JosonChan1998, @lvhan028, @Czm369, @BIGWangYuDong, @RangiLyu, @hhaAndroid, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@lihua199710 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9388\">https://github.com/open-mmlab/mmdetection/pull/9388</a>\n</li>\n<li>@twmht made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9450\">https://github.com/open-mmlab/mmdetection/pull/9450</a>\n</li>\n<li>@tianleiSHI made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9453\">https://github.com/open-mmlab/mmdetection/pull/9453</a>\n</li>\n<li>@kitecats made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9481\">https://github.com/open-mmlab/mmdetection/pull/9481</a>\n</li>\n<li>@QJC123654 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9468\">https://github.com/open-mmlab/mmdetection/pull/9468</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc4...v3.0.0rc5\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc4...v3.0.0rc5</a></p>\n","assets":[]},{"id":84174492,"tag":"v3.0.0rc4","author":"ZwwWayne","name":"MMDetection V3.0.0rc4 Release","draft":false,"prerelease":true,"createdAt":"2022-11-25T11:14:25Z","publishedAt":"2022-11-25T11:20:35Z","markdown":"\r\n\r\n## Highlights\r\n\r\n- Support [CondInst](https://arxiv.org/abs/2003.05664)\r\n- Add `projects/` folder, which will be a place for some experimental models/features.\r\n- Support [SparseInst](https://arxiv.org/abs/2203.12827) in [`projects`](./projects/SparseInst/README.md)\r\n\r\n## New Features\r\n\r\n- Support [CondInst](https://arxiv.org/abs/2003.05664) (#9223)\r\n- Add `projects/` folder, which will be a place for some experimental models/features (#9341)\r\n- Support [SparseInst](https://arxiv.org/abs/2203.12827) in [`projects`](./projects/SparseInst/README.md) (#9377)\r\n\r\n## Bug Fixes\r\n\r\n- Fix `pixel_decoder_type` discrimination in MaskFormer Head. (#9176)\r\n- Fix wrong padding value in cached MixUp (#9259)\r\n- Rename `utils/typing.py` to `utils/typing_utils.py` to fix `collect_env` error (#9265)\r\n- Fix resume arg conflict (#9287)\r\n- Fix the configs of Faster R-CNN with caffe backbone (#9319)\r\n- Fix torchserve and update related documentation (#9343)\r\n- Fix bbox refine bug with sigmooid activation (#9538)\r\n\r\n## Improvements\r\n\r\n- Update the docs of GIoU Loss in README (#8810)\r\n- Handle dataset wrapper in `inference_detector` (#9144)\r\n- Update the type of `counts` in COCO’s compressed RLE (#9274)\r\n- Support saving config file in `print_config` (#9276)\r\n- Update docs about video inference (#9305)\r\n- Update guide about model deployment (#9344)\r\n- Fix doc typos of useful tools (#9177)\r\n- Allow to resume from specific checkpoint in CLI (#9284)\r\n- Update FAQ about windows installation issues of pycocotools (#9292)\r\n\r\n\r\n## Contributors\r\n\r\nA total of 13 developers contributed to this release.\r\n\r\nThanks @JunyaoHu, @sanbuphy, @Czm369, @Daa98, @jbwang1997, @BIGWangYuDong, @JosonChan1998, @lvhan028, @RunningLeon, @RangiLyu, @Daa98, @ZwwWayne, @hhaAndroid\r\n\r\n## New Contributors\r\n* @Daa98 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9274\r\n* @lvhan028 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9344\r\n* @JunyaoHu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9383\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc3...v3.0.0rc4","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2003.05664\" rel=\"nofollow\">CondInst</a>\n</li>\n<li>Add <code>projects/</code> folder, which will be a place for some experimental models/features.</li>\n<li>Support <a href=\"https://arxiv.org/abs/2203.12827\" rel=\"nofollow\">SparseInst</a> in <a href=\"./projects/SparseInst/README.md\"><code>projects</code></a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2003.05664\" rel=\"nofollow\">CondInst</a> (#9223)</li>\n<li>Add <code>projects/</code> folder, which will be a place for some experimental models/features (#9341)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2203.12827\" rel=\"nofollow\">SparseInst</a> in <a href=\"./projects/SparseInst/README.md\"><code>projects</code></a> (#9377)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix <code>pixel_decoder_type</code> discrimination in MaskFormer Head. (#9176)</li>\n<li>Fix wrong padding value in cached MixUp (#9259)</li>\n<li>Rename <code>utils/typing.py</code> to <code>utils/typing_utils.py</code> to fix <code>collect_env</code> error (#9265)</li>\n<li>Fix resume arg conflict (#9287)</li>\n<li>Fix the configs of Faster R-CNN with caffe backbone (#9319)</li>\n<li>Fix torchserve and update related documentation (#9343)</li>\n<li>Fix bbox refine bug with sigmooid activation (#9538)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Update the docs of GIoU Loss in README (#8810)</li>\n<li>Handle dataset wrapper in <code>inference_detector</code> (#9144)</li>\n<li>Update the type of <code>counts</code> in COCO’s compressed RLE (#9274)</li>\n<li>Support saving config file in <code>print_config</code> (#9276)</li>\n<li>Update docs about video inference (#9305)</li>\n<li>Update guide about model deployment (#9344)</li>\n<li>Fix doc typos of useful tools (#9177)</li>\n<li>Allow to resume from specific checkpoint in CLI (#9284)</li>\n<li>Update FAQ about windows installation issues of pycocotools (#9292)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 13 developers contributed to this release.</p>\n<p>Thanks @JunyaoHu, @sanbuphy, @Czm369, @Daa98, @jbwang1997, @BIGWangYuDong, @JosonChan1998, @lvhan028, @RunningLeon, @RangiLyu, @Daa98, @ZwwWayne, @hhaAndroid</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@Daa98 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9274\">https://github.com/open-mmlab/mmdetection/pull/9274</a>\n</li>\n<li>@lvhan028 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9344\">https://github.com/open-mmlab/mmdetection/pull/9344</a>\n</li>\n<li>@JunyaoHu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9383\">https://github.com/open-mmlab/mmdetection/pull/9383</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc3...v3.0.0rc4\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc3...v3.0.0rc4</a></p>\n","assets":[]},{"id":83998095,"tag":"v2.26.0","author":"ZwwWayne","name":"MMDetection V2.26.0 Release","draft":false,"prerelease":false,"createdAt":"2022-11-23T14:23:53Z","publishedAt":"2022-11-23T14:27:06Z","markdown":"## Highlights\r\n\r\n- Support training on [NPU](docs/en/device/npu.md) (#9267)\r\n\r\n## Bug Fixes\r\n\r\n- Fix RPN visualization (#9151)\r\n- Fix readthedocs by freezing the dependency versions (#9154)\r\n- Fix device argument error in MMDet_Tutorial.ipynb (#9112)\r\n- Fix solov2 cannot dealing with empty gt image (#9185)\r\n- Fix random flipping ratio comparison of mixup image (#9336)\r\n\r\n## Improvements\r\n\r\n- Complement necessary argument of seg_suffix of cityscapes (#9330)\r\n- Support copy paste based on bbox when there is no gt mask (#8905)\r\n- Make scipy as a default dependency in runtime (#9186)\r\n\r\n## Documents\r\n\r\n- Delete redundant Chinese characters in docs (#9175)\r\n- Add MMEval in README (#9217)\r\n\r\n## Contributors\r\n\r\nA total of 11 developers contributed to this release.\r\nThanks @wangjiangben-hw, @motokimura, @AdorableJiang, @BainOuO, @JarvisKevin, @wanghonglie, @zytx121, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne\r\n\r\n\r\n## New Contributors\r\n* @JarvisKevin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8905\r\n* @BainOuO made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9175\r\n* @wangjiangben-hw made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9267\r\n* @AdorableJiang made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9330\r\n* @motokimura made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9336\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.25.3...v2.26.0","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support training on <a href=\"docs/en/device/npu.md\">NPU</a> (#9267)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix RPN visualization (#9151)</li>\n<li>Fix readthedocs by freezing the dependency versions (#9154)</li>\n<li>Fix device argument error in MMDet_Tutorial.ipynb (#9112)</li>\n<li>Fix solov2 cannot dealing with empty gt image (#9185)</li>\n<li>Fix random flipping ratio comparison of mixup image (#9336)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Complement necessary argument of seg_suffix of cityscapes (#9330)</li>\n<li>Support copy paste based on bbox when there is no gt mask (#8905)</li>\n<li>Make scipy as a default dependency in runtime (#9186)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Documents</h2><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Delete redundant Chinese characters in docs (#9175)</li>\n<li>Add MMEval in README (#9217)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 11 developers contributed to this release.\nThanks @wangjiangben-hw, @motokimura, @AdorableJiang, @BainOuO, @JarvisKevin, @wanghonglie, @zytx121, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@JarvisKevin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8905\">https://github.com/open-mmlab/mmdetection/pull/8905</a>\n</li>\n<li>@BainOuO made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9175\">https://github.com/open-mmlab/mmdetection/pull/9175</a>\n</li>\n<li>@wangjiangben-hw made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9267\">https://github.com/open-mmlab/mmdetection/pull/9267</a>\n</li>\n<li>@AdorableJiang made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9330\">https://github.com/open-mmlab/mmdetection/pull/9330</a>\n</li>\n<li>@motokimura made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9336\">https://github.com/open-mmlab/mmdetection/pull/9336</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.25.3...v2.26.0\">https://github.com/open-mmlab/mmdetection/compare/v2.25.3...v2.26.0</a></p>\n","assets":[]},{"id":82152008,"tag":"v3.0.0rc3","author":"ZwwWayne","name":"MMDetection V3.0.0rc3 Release","draft":false,"prerelease":true,"createdAt":"2022-11-05T05:36:14Z","publishedAt":"2022-11-05T05:38:41Z","markdown":"## Highlights\r\n\r\n- Support [CrowdDet](https://arxiv.org/abs/2003.09163) and [EIoU Loss](https://ieeexplore.ieee.org/document/9429909)\r\n- Support training detection models in Detectron2\r\n- Refactor Fast R-CNN\r\n- **Note**: In this version, we upgrade the minimum version requirement of MMEngine to 0.3.0 to use `ignore_key` of `ConcatDataset` for training VOC datasets (#9058)\r\n\r\n## New Features\r\n\r\n- Support [CrowdDet](https://arxiv.org/abs/2003.09163) (#8744)\r\n- Support training detection models in Detectron2 with examples of Mask R-CNN, Faster R-CNN, and RetinaNet (#8672)\r\n- Support [EIoU Loss](https://ieeexplore.ieee.org/document/9429909) (#9086)\r\n\r\n## Bug Fixes\r\n\r\n- Fix `XMLDataset` image size error (#9216)\r\n- Fix bugs of empty_instances when predicting without nms in roi_head (#9015)\r\n- Fix the config file of DETR (#9158)\r\n- Fix SOLOv2 cannot dealing with empty gt image (#9192)\r\n- Fix inference demo (#9153)\r\n- Add `ignore_key` in VOC `ConcatDataset` (#9058)\r\n- Fix dumping results issue in test scripts. (#9241)\r\n- Fix configs of training coco subsets on MMDet 3.x (#9225)\r\n- Fix corner2hbox of HorizontalBoxes for supporting empty bboxes (#9140)\r\n\r\n## Improvements\r\n\r\n- Refactor Fast R-CNN (#9132)\r\n- Clean requirements of mmcv-full due to SyncBN (#9207)\r\n- Support training detection models in detectron2 (#8672)\r\n- Add `box_type` support for `DynamicSoftLabelAssigner` (#9179)\r\n- Make scipy as a default dependency in runtime (#9187)\r\n- Update eval_metric (#9062)\r\n- Add `seg_map_suffix` in `BaseDetDataset` (#9088)\r\n\r\n## New Contributors\r\n\r\n- @Wwupup made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9086>\r\n- @sanbuphy made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9153>\r\n- @cxiang26 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9158>\r\n- @JosonChan1998 made their first contribution in <https://github.com/open-mmlab/mmdetection/pull/9225>\r\n\r\n## Contributors\r\n\r\nA total of 13 developers contributed to this release.\r\n\r\nThanks @wanghonglie, @Wwupup, @sanbuphy, @BIGWangYuDong, @liuyanyi, @cxiang26, @jbwang1997, @ZwwWayne, @yuyoujiang, @RangiLyu, @hhaAndroid, @JosonChan1998, @Czm369\r\n\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc2...v3.0.0rc3","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2003.09163\" rel=\"nofollow\">CrowdDet</a> and <a href=\"https://ieeexplore.ieee.org/document/9429909\" rel=\"nofollow\">EIoU Loss</a>\n</li>\n<li>Support training detection models in Detectron2</li>\n<li>Refactor Fast R-CNN</li>\n<li>\n<strong>Note</strong>: In this version, we upgrade the minimum version requirement of MMEngine to 0.3.0 to use <code>ignore_key</code> of <code>ConcatDataset</code> for training VOC datasets (#9058)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2003.09163\" rel=\"nofollow\">CrowdDet</a> (#8744)</li>\n<li>Support training detection models in Detectron2 with examples of Mask R-CNN, Faster R-CNN, and RetinaNet (#8672)</li>\n<li>Support <a href=\"https://ieeexplore.ieee.org/document/9429909\" rel=\"nofollow\">EIoU Loss</a> (#9086)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix <code>XMLDataset</code> image size error (#9216)</li>\n<li>Fix bugs of empty_instances when predicting without nms in roi_head (#9015)</li>\n<li>Fix the config file of DETR (#9158)</li>\n<li>Fix SOLOv2 cannot dealing with empty gt image (#9192)</li>\n<li>Fix inference demo (#9153)</li>\n<li>Add <code>ignore_key</code> in VOC <code>ConcatDataset</code> (#9058)</li>\n<li>Fix dumping results issue in test scripts. (#9241)</li>\n<li>Fix configs of training coco subsets on MMDet 3.x (#9225)</li>\n<li>Fix corner2hbox of HorizontalBoxes for supporting empty bboxes (#9140)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Refactor Fast R-CNN (#9132)</li>\n<li>Clean requirements of mmcv-full due to SyncBN (#9207)</li>\n<li>Support training detection models in detectron2 (#8672)</li>\n<li>Add <code>box_type</code> support for <code>DynamicSoftLabelAssigner</code> (#9179)</li>\n<li>Make scipy as a default dependency in runtime (#9187)</li>\n<li>Update eval_metric (#9062)</li>\n<li>Add <code>seg_map_suffix</code> in <code>BaseDetDataset</code> (#9088)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@Wwupup made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9086\">https://github.com/open-mmlab/mmdetection/pull/9086</a>\n</li>\n<li>@sanbuphy made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9153\">https://github.com/open-mmlab/mmdetection/pull/9153</a>\n</li>\n<li>@cxiang26 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9158\">https://github.com/open-mmlab/mmdetection/pull/9158</a>\n</li>\n<li>@JosonChan1998 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9225\">https://github.com/open-mmlab/mmdetection/pull/9225</a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 13 developers contributed to this release.</p>\n<p>Thanks @wanghonglie, @Wwupup, @sanbuphy, @BIGWangYuDong, @liuyanyi, @cxiang26, @jbwang1997, @ZwwWayne, @yuyoujiang, @RangiLyu, @hhaAndroid, @JosonChan1998, @Czm369</p>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc2...v3.0.0rc3\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc2...v3.0.0rc3</a></p>\n","assets":[]},{"id":80890557,"tag":"v2.25.3","author":"RangiLyu","name":"MMDetection V2.25.3 Release","draft":false,"prerelease":false,"createdAt":"2022-10-25T08:06:32Z","publishedAt":"2022-10-25T08:18:51Z","markdown":"### Bug Fixes\r\n\r\n- Skip remote sync when wandb is offline (#8755)\r\n- Fix jpg to png bug when using seg maps (#9078)\r\n\r\n### Improvements\r\n\r\n- Fix typo in warning (#8844)\r\n- Fix CI for timm, pycocotools, onnx (#9034)\r\n- Upgrade pre-commit hooks (#8964)\r\n\r\n### Documents\r\n\r\n- Update BoundedIoULoss config in readme (#8808)\r\n- Fix Faster R-CNN Readme (#8803)\r\n- Update location of test_cfg and train_cfg (#8792)\r\n- Fix issue template (#8966)\r\n- Update random sampler docstring (#9033)\r\n- Fix wrong image link (#9054)\r\n- Fix FPG readme (#9041)\r\n\r\n### Contributors\r\n\r\nA total of 13 developers contributed to this release.\r\nThanks @Zheng-LinXiao, @i-aki-y, @fbagci, @sudoAimer, @Czm369, @DrRyanHuang, @RangiLyu, @wanghonglie, @shinya7y, @Ryoo72, @akshaygulabrao, @gy-7, @Neesky\r\n\r\n\r\n## New Contributors\r\n* @i-aki-y made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8755\r\n* @sudoAimer made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8803\r\n* @DrRyanHuang made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8792\r\n* @Ryoo72 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9033\r\n* @akshaygulabrao made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9054\r\n* @gy-7 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9041\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.25.2...v2.25.3","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Skip remote sync when wandb is offline (#8755)</li>\n<li>Fix jpg to png bug when using seg maps (#9078)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix typo in warning (#8844)</li>\n<li>Fix CI for timm, pycocotools, onnx (#9034)</li>\n<li>Upgrade pre-commit hooks (#8964)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Documents</h3><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Update BoundedIoULoss config in readme (#8808)</li>\n<li>Fix Faster R-CNN Readme (#8803)</li>\n<li>Update location of test_cfg and train_cfg (#8792)</li>\n<li>Fix issue template (#8966)</li>\n<li>Update random sampler docstring (#9033)</li>\n<li>Fix wrong image link (#9054)</li>\n<li>Fix FPG readme (#9041)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 13 developers contributed to this release.\nThanks @Zheng-LinXiao, @i-aki-y, @fbagci, @sudoAimer, @Czm369, @DrRyanHuang, @RangiLyu, @wanghonglie, @shinya7y, @Ryoo72, @akshaygulabrao, @gy-7, @Neesky</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@i-aki-y made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8755\">https://github.com/open-mmlab/mmdetection/pull/8755</a>\n</li>\n<li>@sudoAimer made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8803\">https://github.com/open-mmlab/mmdetection/pull/8803</a>\n</li>\n<li>@DrRyanHuang made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8792\">https://github.com/open-mmlab/mmdetection/pull/8792</a>\n</li>\n<li>@Ryoo72 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9033\">https://github.com/open-mmlab/mmdetection/pull/9033</a>\n</li>\n<li>@akshaygulabrao made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9054\">https://github.com/open-mmlab/mmdetection/pull/9054</a>\n</li>\n<li>@gy-7 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9041\">https://github.com/open-mmlab/mmdetection/pull/9041</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.25.2...v2.25.3\">https://github.com/open-mmlab/mmdetection/compare/v2.25.2...v2.25.3</a></p>\n","assets":[]},{"id":80585626,"tag":"v3.0.0rc2","author":"ZwwWayne","name":"MMDetection V3.0.0rc2 Release","draft":false,"prerelease":true,"createdAt":"2022-10-21T10:20:02Z","publishedAt":"2022-10-21T10:23:17Z","markdown":"## Highlights\r\n\r\n- Support [imagenet pre-training](configs/rtmdet/cspnext_imagenet_pretrain) for RTMDet's backbone\r\n\r\n## New Features\r\n\r\n- Support [imagenet pre-training](configs/rtmdet/cspnext_imagenet_pretrain) for RTMDet's backbone (#8887)\r\n- Add `CrowdHumanDataset` and Metric (#8430)\r\n- Add `FixShapeResize` to support resize of fixed shape (#8665)\r\n\r\n## Bug Fixes\r\n\r\n- Fix `ConcatDataset` Import Error (#8909)\r\n- Fix `CircleCI` and `readthedoc` build failed (#8980, #8963)\r\n- Fix bitmap mask translate when `out_shape` is different (#8993)\r\n- Fix inconsistency in `Conv2d` weight channels (#8948)\r\n- Fix bugs when plotting loss curve by analyze_logs.py (#8944)\r\n- Fix type change of labels in `albumentations` (#9074)\r\n- Fix some docs and types error (#8818)\r\n- Update memory occupation of `RTMDet` in metafile (#9098)\r\n- Fix wrong arguments of `OpenImageMetrics` in the config (#9061)\r\n\r\n## Improvements\r\n\r\n- Refactor standard roi head with `box type` (#8658)\r\n- Support mask concatenation in `BitmapMasks` and `PolygonMasks` (#9006)\r\n- Update PyTorch and dependencies' version in dockerfile (#8845)\r\n- Update `robustness_eval.py` and `print_config` (#8452)\r\n- Make compatible with `ConfigDict` and `dict` in `dense_heads` (#8942)\r\n- Support logging coco metric copypaste (#9012)\r\n- Remove `Normalize` transform (#8913)\r\n- Support jittering the color of different instances of the same class (#8988)\r\n- Add assertion for missing key in `PackDetInputs` (#8982)\r\n\r\n## Contributors\r\n\r\nA total of 13 developers contributed to this release.\r\n\r\nThanks @RangiLyu, @jbwang1997, @wanghonglie, @Chan-Sun, @RangeKing, @chhluo, @MambaWong, @yuyoujiang, @hhaAndroid, @sltlls, @Nioolek, @ZwwWayne, @wufan-tb\r\n\r\n## New Contributors\r\n* @Chan-Sun made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8909\r\n* @MambaWong made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8913\r\n* @yuyoujiang made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8437\r\n* @sltlls made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8944\r\n* @Nioolek made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8845\r\n* @wufan-tb made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9061\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc1...v3.0.0rc2","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"configs/rtmdet/cspnext_imagenet_pretrain\">imagenet pre-training</a> for RTMDet's backbone</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"configs/rtmdet/cspnext_imagenet_pretrain\">imagenet pre-training</a> for RTMDet's backbone (#8887)</li>\n<li>Add <code>CrowdHumanDataset</code> and Metric (#8430)</li>\n<li>Add <code>FixShapeResize</code> to support resize of fixed shape (#8665)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix <code>ConcatDataset</code> Import Error (#8909)</li>\n<li>Fix <code>CircleCI</code> and <code>readthedoc</code> build failed (#8980, #8963)</li>\n<li>Fix bitmap mask translate when <code>out_shape</code> is different (#8993)</li>\n<li>Fix inconsistency in <code>Conv2d</code> weight channels (#8948)</li>\n<li>Fix bugs when plotting loss curve by analyze_logs.py (#8944)</li>\n<li>Fix type change of labels in <code>albumentations</code> (#9074)</li>\n<li>Fix some docs and types error (#8818)</li>\n<li>Update memory occupation of <code>RTMDet</code> in metafile (#9098)</li>\n<li>Fix wrong arguments of <code>OpenImageMetrics</code> in the config (#9061)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Refactor standard roi head with <code>box type</code> (#8658)</li>\n<li>Support mask concatenation in <code>BitmapMasks</code> and <code>PolygonMasks</code> (#9006)</li>\n<li>Update PyTorch and dependencies' version in dockerfile (#8845)</li>\n<li>Update <code>robustness_eval.py</code> and <code>print_config</code> (#8452)</li>\n<li>Make compatible with <code>ConfigDict</code> and <code>dict</code> in <code>dense_heads</code> (#8942)</li>\n<li>Support logging coco metric copypaste (#9012)</li>\n<li>Remove <code>Normalize</code> transform (#8913)</li>\n<li>Support jittering the color of different instances of the same class (#8988)</li>\n<li>Add assertion for missing key in <code>PackDetInputs</code> (#8982)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 13 developers contributed to this release.</p>\n<p>Thanks @RangiLyu, @jbwang1997, @wanghonglie, @Chan-Sun, @RangeKing, @chhluo, @MambaWong, @yuyoujiang, @hhaAndroid, @sltlls, @Nioolek, @ZwwWayne, @wufan-tb</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@Chan-Sun made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8909\">https://github.com/open-mmlab/mmdetection/pull/8909</a>\n</li>\n<li>@MambaWong made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8913\">https://github.com/open-mmlab/mmdetection/pull/8913</a>\n</li>\n<li>@yuyoujiang made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8437\">https://github.com/open-mmlab/mmdetection/pull/8437</a>\n</li>\n<li>@sltlls made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8944\">https://github.com/open-mmlab/mmdetection/pull/8944</a>\n</li>\n<li>@Nioolek made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8845\">https://github.com/open-mmlab/mmdetection/pull/8845</a>\n</li>\n<li>@wufan-tb made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/9061\">https://github.com/open-mmlab/mmdetection/pull/9061</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc1...v3.0.0rc2\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc1...v3.0.0rc2</a></p>\n","assets":[]},{"id":78108470,"tag":"v3.0.0rc1","author":"ZwwWayne","name":"MMDetection V3.0.0rc1 Release","draft":false,"prerelease":true,"createdAt":"2022-09-26T08:02:31Z","publishedAt":"2022-09-26T08:05:44Z","markdown":"### Highlights\r\n\r\n- Release a high-precision, low-latency single-stage object detector [RTMDet](configs/rtmdet).\r\n\r\n### Bug Fixes\r\n\r\n- Fix UT to be compatible with PyTorch 1.6 (#8707)\r\n- Fix `NumClassCheckHook` bug when model is wrapped (#8794)\r\n- Update the right URL of R-50-FPN with BoundedIoULoss (#8805)\r\n- Fix potential bug of indices in RandAugment (#8826)\r\n- Fix some types and links (#8839, #8820, #8793, #8868)\r\n- Fix incorrect background fill values in `FSAF` and `RepPoints` Head (#8813)\r\n\r\n### Improvements\r\n\r\n- Refactored anchor head and base head with `box type` (#8625)\r\n- Refactored `SemiBaseDetector` and `SoftTeacher` (#8786)\r\n- Add list to dict keys to avoid modify loss dict (#8828)\r\n- Update `analyze_results.py` , `analyze_logs.py` and `loading.py` (#8430, #8402, #8784)\r\n- Support dump results in `test.py` (#8814)\r\n- Check empty predictions in `DetLocalVisualizer._draw_instances` (#8830)\r\n- Fix `floordiv` warning in `SOLO` (#8738)\r\n\r\n### Contributors\r\n\r\nA total of 16 developers contributed to this release.\r\n\r\nThanks @ZwwWayne, @jbwang1997, @Czm369, @ice-tong, @Zheng-LinXiao, @chhluo, @RangiLyu, @liuyanyi, @wanghonglie, @levan92, @JiayuXu0, @nye0, @hhaAndroid, @xin-li-67, @shuxp, @zytx121\r\n\r\n## New Contributors\r\n* @YvesXu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8871\r\n* @xin-li-67 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8868\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc0...v3.0.0rc1","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Release a high-precision, low-latency single-stage object detector <a href=\"configs/rtmdet\">RTMDet</a>.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix UT to be compatible with PyTorch 1.6 (#8707)</li>\n<li>Fix <code>NumClassCheckHook</code> bug when model is wrapped (#8794)</li>\n<li>Update the right URL of R-50-FPN with BoundedIoULoss (#8805)</li>\n<li>Fix potential bug of indices in RandAugment (#8826)</li>\n<li>Fix some types and links (#8839, #8820, #8793, #8868)</li>\n<li>Fix incorrect background fill values in <code>FSAF</code> and <code>RepPoints</code> Head (#8813)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Refactored anchor head and base head with <code>box type</code> (#8625)</li>\n<li>Refactored <code>SemiBaseDetector</code> and <code>SoftTeacher</code> (#8786)</li>\n<li>Add list to dict keys to avoid modify loss dict (#8828)</li>\n<li>Update <code>analyze_results.py</code> , <code>analyze_logs.py</code> and <code>loading.py</code> (#8430, #8402, #8784)</li>\n<li>Support dump results in <code>test.py</code> (#8814)</li>\n<li>Check empty predictions in <code>DetLocalVisualizer._draw_instances</code> (#8830)</li>\n<li>Fix <code>floordiv</code> warning in <code>SOLO</code> (#8738)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 16 developers contributed to this release.</p>\n<p>Thanks @ZwwWayne, @jbwang1997, @Czm369, @ice-tong, @Zheng-LinXiao, @chhluo, @RangiLyu, @liuyanyi, @wanghonglie, @levan92, @JiayuXu0, @nye0, @hhaAndroid, @xin-li-67, @shuxp, @zytx121</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@YvesXu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8871\">https://github.com/open-mmlab/mmdetection/pull/8871</a>\n</li>\n<li>@xin-li-67 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8868\">https://github.com/open-mmlab/mmdetection/pull/8868</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc0...v3.0.0rc1\">https://github.com/open-mmlab/mmdetection/compare/v3.0.0rc0...v3.0.0rc1</a></p>\n","assets":[]},{"id":77370776,"tag":"v2.25.2","author":"ZwwWayne","name":"MMDetection V2.25.2 Release","draft":false,"prerelease":false,"createdAt":"2022-09-16T08:23:46Z","publishedAt":"2022-09-16T08:27:32Z","markdown":"### Bug Fixes\r\n\r\n- Fix DyDCNv2 RuntimeError (#8485)\r\n- Fix repeated import of CascadeRPNHead (#8578)\r\n- Fix absolute positional embedding of swin backbone (#8127)\r\n- Fix get train_pipeline method of val workflow (#8575)\r\n\r\n### Improvements\r\n\r\n- Upgrade onnxsim to at least 0.4.0 (#8383)\r\n- Support tuple format in analyze_results script (#8549)\r\n- Fix floordiv warning (#8648)\r\n\r\n### Documents\r\n\r\n- Fix typo in HTC link (#8487)\r\n- Fix docstring of `BboxOverlaps2D` (#8512)\r\n- Added missed Chinese tutorial link (#8564)\r\n- Fix mistakes in gaussian radius formula (#8607)\r\n- Update config documentation about how to Add WandB Hook (#8663)\r\n- Add mmengine link in readme (#8799)\r\n- Update issue template (#8802)\r\n\r\n### Ongoing changes\r\n\r\n- Support training YOLOv3 on IPU #8552 \r\n- Support RF-NeXt #8191  \r\n\r\n### Contributors\r\n\r\nA total of 16 developers contributed to this release.\r\nThanks @daquexian, @lyq10085, @ZwwWayne, @fbagci, @BubblyYi, @fathomson, @ShunchiZhang, @ceasona, @Happylkx, @normster, @chhluo, @Lehsuby, @JiayuXu0, @Nourollah, @hewanru-bit, @RangiLyu\r\n\r\n## New Contributors\r\n* @RockeyCoss made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8733\r\n* @hewanru-bit made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8799\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.25.1...v2.25.2","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix DyDCNv2 RuntimeError (#8485)</li>\n<li>Fix repeated import of CascadeRPNHead (#8578)</li>\n<li>Fix absolute positional embedding of swin backbone (#8127)</li>\n<li>Fix get train_pipeline method of val workflow (#8575)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Upgrade onnxsim to at least 0.4.0 (#8383)</li>\n<li>Support tuple format in analyze_results script (#8549)</li>\n<li>Fix floordiv warning (#8648)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Documents</h3><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix typo in HTC link (#8487)</li>\n<li>Fix docstring of <code>BboxOverlaps2D</code> (#8512)</li>\n<li>Added missed Chinese tutorial link (#8564)</li>\n<li>Fix mistakes in gaussian radius formula (#8607)</li>\n<li>Update config documentation about how to Add WandB Hook (#8663)</li>\n<li>Add mmengine link in readme (#8799)</li>\n<li>Update issue template (#8802)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Ongoing changes</h3><a id=\"user-content-ongoing-changes\" class=\"anchor\" aria-label=\"Permalink: Ongoing changes\" href=\"#ongoing-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support training YOLOv3 on IPU #8552</li>\n<li>Support RF-NeXt #8191</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 16 developers contributed to this release.\nThanks @daquexian, @lyq10085, @ZwwWayne, @fbagci, @BubblyYi, @fathomson, @ShunchiZhang, @ceasona, @Happylkx, @normster, @chhluo, @Lehsuby, @JiayuXu0, @Nourollah, @hewanru-bit, @RangiLyu</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@RockeyCoss made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8733\">https://github.com/open-mmlab/mmdetection/pull/8733</a>\n</li>\n<li>@hewanru-bit made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8799\">https://github.com/open-mmlab/mmdetection/pull/8799</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.25.1...v2.25.2\">https://github.com/open-mmlab/mmdetection/compare/v2.25.1...v2.25.2</a></p>\n","assets":[]},{"id":75976157,"tag":"v3.0.0rc0","author":"ZwwWayne","name":"MMDetection V3.0.0rc0 Release","draft":false,"prerelease":true,"createdAt":"2022-09-01T03:57:14Z","publishedAt":"2022-09-01T03:57:50Z","markdown":"We are excited to announce the release of MMDetection 3.0.0rc0. MMDet 3.0.0rc0 is the first version of MMDetection 3.x, a part of the OpenMMLab 2.0 projects. Built upon the new [training engine](https://github.com/open-mmlab/mmengine), MMDet 3.x unifies the interfaces of the dataset, models, evaluation, and visualization with faster training and testing speed. It also provides a general semi-supervised object detection framework and strong baselines.\r\n\r\n## Highlights\r\n\r\n1. **New engine**. MMDet 3.x is based on [MMEngine](https://github.com/open-mmlab/mmengine), which provides a universal and powerful runner that allows more flexible customizations and significantly simplifies the entry points of high-level interfaces.\r\n\r\n2. **Unified interfaces**. As a part of the OpenMMLab 2.0 projects, MMDet 3.x unifies and refactors the interfaces and internal logic of training, testing, datasets, models, evaluation, and visualization. All the OpenMMLab 2.0 projects share the same design in those interfaces and logic to allow the emergence of multi-task/modality algorithms.\r\n\r\n3. **Faster speed**. We optimize the training and inference speed for common models and configurations, achieving faster or similar speed in comparison with [Detection2](https://github.com/facebookresearch/detectron2/). Model details of benchmark will be updated in [this note](./benchmark.md#comparison-with-detectron2).\r\n\r\n4. **General semi-supervised object detection**. Benefitting from the unified interfaces, we support a general semi-supervised learning framework that works with all the object detectors supported in MMDet 3.x. Please refer to [semi-supervised object detection](../user_guides/semi_det.md) for details.\r\n\r\n5. **Strong baselines**. We release strong baselines of many popular models to enable fair comparisons among state-of-the-art models.\r\n\r\n6. **New features and algorithms**:\r\n\r\n   - Enable all the single-stage detectors to serve as region proposal networks\r\n   - [SoftTeacher](https://arxiv.org/abs/2106.09018)\r\n   - [the updated CenterNet](https://arxiv.org/abs/2103.07461)\r\n\r\n7. **More documentation and tutorials**. We add a bunch of documentation and tutorials to help users get started more smoothly. Read it [here](https://mmdetection.readthedocs.io/en/3.x/).\r\n\r\n## Breaking Changes\r\n\r\nMMDet 3.x has gone through big changes to have better design, higher efficiency, more flexibility, and more unified interfaces.\r\nBesides the changes in API, we briefly list the major breaking changes in this section.\r\nWe will update the [migration guide](../migration.md) to provide complete details and migration instructions.\r\nUsers can also refer to the [API doc](https://mmdetection.readthedocs.io/en/3.x/) for more details.\r\n\r\n### Dependencies\r\n\r\n- MMDet 3.x runs on PyTorch>=1.6. We have deprecated the support of PyTorch 1.5 to embrace mixed precision training and other new features since PyTorch 1.6. Some models can still run on PyTorch 1.5, but the full functionality of MMDet 3.x is not guaranteed.\r\n- MMDet 3.x relies on MMEngine to run. MMEngine is a new foundational library for training deep learning models of OpenMMLab and is the core dependency of OpenMMLab 2.0 projects. The dependencies of file IO and training are migrated from MMCV 1.x to MMEngine.\r\n- MMDet 3.x relies on MMCV>=2.0.0rc0. Although MMCV no longer maintains the training functionalities since 2.0.0rc0, MMDet 3.x relies on the data transforms, CUDA operators, and image processing interfaces in MMCV. Note that the package `mmcv` is the version that provides pre-built CUDA operators and `mmcv-lite` does not since MMCV 2.0.0rc0, while `mmcv-full` has been deprecated since 2.0.0rc0.\r\n\r\n### Training and testing\r\n\r\n- MMDet 3.x uses Runner in [MMEngine](https://github.com/open-mmlab/mmengine) rather than that in MMCV. The new Runner implements and unifies the building logic of the dataset, model, evaluation, and visualizer. Therefore, MMDet 3.x no longer maintains the building logic of those modules in `mmdet.train.apis` and `tools/train.py`. Those codes have been migrated into [MMEngine](https://github.com/open-mmlab/mmengine/blob/main/mmengine/runner/runner.py). Please refer to the [migration guide of Runner in MMEngine](https://mmengine.readthedocs.io/en/latest/migration/runner.html) for more details.\r\n- The Runner in MMEngine also supports testing and validation. The testing scripts are also simplified, which has similar logic to that in training scripts to build the runner.\r\n- The execution points of hooks in the new Runner have been enriched to allow more flexible customization. Please refer to the [migration guide of Hook in MMEngine](https://mmengine.readthedocs.io/en/latest/migration/hook.html) for more details.\r\n- Learning rate and momentum schedules have been migrated from Hook to [Parameter Scheduler in MMEngine](https://mmengine.readthedocs.io/en/latest/tutorials/param_scheduler.html). Please refer to the [migration guide of Parameter Scheduler in MMEngine](https://mmengine.readthedocs.io/en/latest/migration/param_scheduler.html) for more details.\r\n\r\n### Configs\r\n\r\n- The [Runner in MMEngine](https://github.com/open-mmlab/mmengine/blob/main/mmengine/runner/runner.py) uses a different config structure to ease the understanding of the components in the runner. Users can read the [config example of MMDet 3.x](../user_guides/config.md) or refer to the [migration guide in MMEngine](https://mmengine.readthedocs.io/en/latest/migration/runner.html) for migration details.\r\n- The file names of configs and models are also refactored to follow the new rules unified across OpenMMLab 2.0 projects. The names of checkpoints are not updated for now as there is no BC-breaking of model weights between MMDet 3.x and 2.x. We will progressively replace all the model weights with those trained in MMDet 3.x. Please refer to the [user guides of config](../user_guides/config.md) for more details.\r\n\r\n### Dataset\r\n\r\nThe Dataset classes implemented in MMDet 3.x all inherit from the `BaseDetDataset`, which inherits from the [BaseDataset in MMEngine](https://mmengine.readthedocs.io/en/latest/advanced_tutorials/basedataset.html). In addition to the changes in interfaces, there are several changes in Dataset in MMDet 3.x.\r\n\r\n- All the datasets support serializing the internal data list to reduce the memory when multiple workers are built for data loading.\r\n- The internal data structure in the dataset is changed to be self-contained (without losing information like class names in MMDet 2.x) while keeping simplicity.\r\n- The evaluation functionality of each dataset has been removed from the dataset so that some specific evaluation metrics like COCO AP can be used to evaluate the prediction on other datasets.\r\n\r\n### Data Transforms\r\n\r\nThe data transforms in MMDet 3.x all inherits from `BaseTransform` in MMCV>=2.0.0rc0, which defines a new convention in OpenMMLab 2.0 projects.\r\nBesides the interface changes, there are several changes listed below:\r\n\r\n- The functionality of some data transforms (e.g., `Resize`) are decomposed into several transforms to simplify and clarify the usages.\r\n- The format of data dict processed by each data transform is changed according to the new data structure of dataset.\r\n- Some inefficient data transforms (e.g., normalization and padding) are moved into the data preprocessor of the model to improve data loading and training speed.\r\n- The same data transforms in different OpenMMLab 2.0 libraries have the same augmentation implementation and the logic given the same arguments, i.e., `Resize` in MMDet 3.x and MMSeg 1.x will resize the image in the exact same manner given the same arguments.\r\n\r\n### Model\r\n\r\nThe models in MMDet 3.x all inherit from `BaseModel` in MMEngine, which defines a new convention of models in OpenMMLab 2.0 projects.\r\nUsers can refer to [the tutorial of the model in MMengine](https://mmengine.readthedocs.io/en/latest/tutorials/model.html) for more details.\r\nAccordingly, there are several changes as the following:\r\n\r\n- The model interfaces, including the input and output formats, are significantly simplified and unified following the new convention in MMDet 3.x.\r\n  Specifically, all the input data in training and testing are packed into `inputs` and `data_samples`, where `inputs` contains model inputs like a list of image tensors, and `data_samples` contains other information of the current data sample such as ground truths, region proposals, and model predictions. In this way, different tasks in MMDet 3.x can share the same input arguments, which makes the models more general and suitable for multi-task learning and some flexible training paradigms like semi-supervised learning.\r\n- The model has a data preprocessor module, which is used to pre-process the input data of the model. In MMDet 3.x, the data preprocessor usually does the necessary steps to form the input images into a batch, such as padding. It can also serve as a place for some special data augmentations or more efficient data transformations like normalization.\r\n- The internal logic of the model has been changed. In MMdet 2.x, model uses `forward_train`, `forward_test`, `simple_test`, and `aug_test` to deal with different model forward logics. In MMDet 3.x and OpenMMLab 2.0, the forward function has three modes: 'loss', 'predict', and 'tensor' for training, inference, and tracing or other purposes, respectively.\r\n  The forward function calls `self.loss`, `self.predict`, and `self._forward` given the modes 'loss', 'predict', and 'tensor', respectively.\r\n\r\n### Evaluation\r\n\r\nThe evaluation in MMDet 2.x strictly binds with the dataset. In contrast, MMDet 3.x decomposes the evaluation from the dataset so that all the detection datasets can evaluate with COCO AP and other metrics implemented in MMDet 3.x.\r\nMMDet 3.x mainly implements corresponding metrics for each dataset, which are manipulated by [Evaluator](https://mmengine.readthedocs.io/en/latest/design/evaluator.html) to complete the evaluation.\r\nUsers can build an evaluator in MMDet 3.x to conduct offline evaluation, i.e., evaluate predictions that may not produce in MMDet 3.x with the dataset as long as the dataset and the prediction follow the dataset conventions. More details can be found in the [tutorial in mmengine](https://mmengine.readthedocs.io/en/latest/tutorials/evaluation.html).\r\n\r\n### Visualization\r\n\r\nThe functions of visualization in MMDet 2.x are removed. Instead, in OpenMMLab 2.0 projects, we use [Visualizer](https://mmengine.readthedocs.io/en/latest/design/visualization.html) to visualize data. MMDet 3.x implements `DetLocalVisualizer` to allow visualization of ground truths, model predictions, feature maps, etc., at any place. It also supports sending the visualization data to any external visualization backends such as Tensorboard.\r\n\r\n## Improvements\r\n\r\n- Optimized training and testing speed of FCOS, RetinaNet, Faster R-CNN, Mask R-CNN, and Cascade R-CNN. The training speed of those models with some common training strategies is also optimized, including those with synchronized batch normalization and mixed precision training.\r\n- Support mixed precision training of all the models. However, some models may get Nan results due to some numerical issues. We will update the documentation and list the results (accuracy of failure) of mixed precision training.\r\n- Release strong baselines of some popular object detectors. Their accuracy and pre-trained checkpoints will be released.\r\n\r\n## Bug Fixes\r\n\r\n- DeepFashion dataset: the config and results have been updated.\r\n\r\n## New Features\r\n\r\n1. Support a general semi-supervised learning framework that works with all the object detectors supported in MMDet 3.x. Please refer to [semi-supervised object detection](../user_guides/semi_det.md) for details.\r\n2. Enable all the single-stage detectors to serve as region proposal networks. We give [an example of using FCOS as RPN](../user_guides/single_stage_as_rpn.md).\r\n3. Support a semi-supervised object detection algorithm: [SoftTeacher](https://arxiv.org/abs/2106.09018).\r\n4. Support [the updated CenterNet](https://arxiv.org/abs/2103.07461).\r\n5. Support data structures `HorizontalBoxes` and `BaseBoxes` to encapsulate different kinds of bounding boxes. We are migrating to use data structures of boxes to replace the use of pure tensor boxes. This will unify the usages of different kinds of bounding boxes in MMDet 3.x and MMRotate 1.x to simplify the implementation and reduce redundant codes.\r\n\r\n## Planned changes\r\n\r\nWe list several planned changes of MMDet 3.0.0rc0 so that the community could more comprehensively know the progress of MMDet 3.x. Feel free to create a PR, issue, or discussion if you are interested, have any suggestions and feedback, or want to participate.\r\n\r\n1. Test-time augmentation: which is supported in MMDet 2.x, is not implemented in this version due to the limited time slot. We will support it in the following releases with a new and simplified design.\r\n2. Inference interfaces: unified inference interfaces will be supported in the future to ease the use of released models.\r\n3. Interfaces of useful tools that can be used in Jupyter Notebook or Colab: more useful tools that are implemented in the `tools` directory will have their python interfaces so that they can be used in Jupyter Notebook, Colab, and downstream libraries.\r\n4. Documentation: we will add more design docs, tutorials, and migration guidance so that the community can deep dive into our new design, participate the future development, and smoothly migrate downstream libraries to MMDet 3.x.\r\n5. Wandb visualization: MMDet 2.x supports data visualization since v2.25.0, which has not been migrated to MMDet 3.x for now. Since WandB provides strong visualization and experiment management capabilities, a `DetWandbVisualizer` and maybe a hook are planned to fully migrate those functionalities from MMDet 2.x.\r\n6. Full support of WiderFace dataset (#8508) and Fast R-CNN: we are verifying their functionalities and will fix related issues soon.\r\n7. Migrate DETR-series algorithms (#8655, #8533) and YOLOv3 on IPU (#8552) from MMDet 2.x.\r\n\r\n## Contributors\r\n\r\nA total of 11 developers contributed to this release.\r\nThanks @shuxp, @wanghonglie, @Czm369, @BIGWangYuDong, @zytx121, @jbwang1997, @chhluo, @jshilong, @RangiLyu, @hhaAndroid, @ZwwWayne\r\n\r\n## New Contributors\r\n* @shuxp made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8612\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v3.0.0rc0","html":"<p>We are excited to announce the release of MMDetection 3.0.0rc0. MMDet 3.0.0rc0 is the first version of MMDetection 3.x, a part of the OpenMMLab 2.0 projects. Built upon the new <a href=\"https://github.com/open-mmlab/mmengine\">training engine</a>, MMDet 3.x unifies the interfaces of the dataset, models, evaluation, and visualization with faster training and testing speed. It also provides a general semi-supervised object detection framework and strong baselines.</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ol>\n<li>\n<p><strong>New engine</strong>. MMDet 3.x is based on <a href=\"https://github.com/open-mmlab/mmengine\">MMEngine</a>, which provides a universal and powerful runner that allows more flexible customizations and significantly simplifies the entry points of high-level interfaces.</p>\n</li>\n<li>\n<p><strong>Unified interfaces</strong>. As a part of the OpenMMLab 2.0 projects, MMDet 3.x unifies and refactors the interfaces and internal logic of training, testing, datasets, models, evaluation, and visualization. All the OpenMMLab 2.0 projects share the same design in those interfaces and logic to allow the emergence of multi-task/modality algorithms.</p>\n</li>\n<li>\n<p><strong>Faster speed</strong>. We optimize the training and inference speed for common models and configurations, achieving faster or similar speed in comparison with <a href=\"https://github.com/facebookresearch/detectron2/\">Detection2</a>. Model details of benchmark will be updated in <a href=\"./benchmark.md#comparison-with-detectron2\">this note</a>.</p>\n</li>\n<li>\n<p><strong>General semi-supervised object detection</strong>. Benefitting from the unified interfaces, we support a general semi-supervised learning framework that works with all the object detectors supported in MMDet 3.x. Please refer to <a href=\"../user_guides/semi_det.md\">semi-supervised object detection</a> for details.</p>\n</li>\n<li>\n<p><strong>Strong baselines</strong>. We release strong baselines of many popular models to enable fair comparisons among state-of-the-art models.</p>\n</li>\n<li>\n<p><strong>New features and algorithms</strong>:</p>\n<ul>\n<li>Enable all the single-stage detectors to serve as region proposal networks</li>\n<li><a href=\"https://arxiv.org/abs/2106.09018\" rel=\"nofollow\">SoftTeacher</a></li>\n<li><a href=\"https://arxiv.org/abs/2103.07461\" rel=\"nofollow\">the updated CenterNet</a></li>\n</ul>\n</li>\n<li>\n<p><strong>More documentation and tutorials</strong>. We add a bunch of documentation and tutorials to help users get started more smoothly. Read it <a href=\"https://mmdetection.readthedocs.io/en/3.x/\" rel=\"nofollow\">here</a>.</p>\n</li>\n</ol>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Breaking Changes</h2><a id=\"user-content-breaking-changes\" class=\"anchor\" aria-label=\"Permalink: Breaking Changes\" href=\"#breaking-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>MMDet 3.x has gone through big changes to have better design, higher efficiency, more flexibility, and more unified interfaces.\nBesides the changes in API, we briefly list the major breaking changes in this section.\nWe will update the <a href=\"../migration.md\">migration guide</a> to provide complete details and migration instructions.\nUsers can also refer to the <a href=\"https://mmdetection.readthedocs.io/en/3.x/\" rel=\"nofollow\">API doc</a> for more details.</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Dependencies</h3><a id=\"user-content-dependencies\" class=\"anchor\" aria-label=\"Permalink: Dependencies\" href=\"#dependencies\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>MMDet 3.x runs on PyTorch&gt;=1.6. We have deprecated the support of PyTorch 1.5 to embrace mixed precision training and other new features since PyTorch 1.6. Some models can still run on PyTorch 1.5, but the full functionality of MMDet 3.x is not guaranteed.</li>\n<li>MMDet 3.x relies on MMEngine to run. MMEngine is a new foundational library for training deep learning models of OpenMMLab and is the core dependency of OpenMMLab 2.0 projects. The dependencies of file IO and training are migrated from MMCV 1.x to MMEngine.</li>\n<li>MMDet 3.x relies on MMCV&gt;=2.0.0rc0. Although MMCV no longer maintains the training functionalities since 2.0.0rc0, MMDet 3.x relies on the data transforms, CUDA operators, and image processing interfaces in MMCV. Note that the package <code>mmcv</code> is the version that provides pre-built CUDA operators and <code>mmcv-lite</code> does not since MMCV 2.0.0rc0, while <code>mmcv-full</code> has been deprecated since 2.0.0rc0.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Training and testing</h3><a id=\"user-content-training-and-testing\" class=\"anchor\" aria-label=\"Permalink: Training and testing\" href=\"#training-and-testing\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>MMDet 3.x uses Runner in <a href=\"https://github.com/open-mmlab/mmengine\">MMEngine</a> rather than that in MMCV. The new Runner implements and unifies the building logic of the dataset, model, evaluation, and visualizer. Therefore, MMDet 3.x no longer maintains the building logic of those modules in <code>mmdet.train.apis</code> and <code>tools/train.py</code>. Those codes have been migrated into <a href=\"https://github.com/open-mmlab/mmengine/blob/main/mmengine/runner/runner.py\">MMEngine</a>. Please refer to the <a href=\"https://mmengine.readthedocs.io/en/latest/migration/runner.html\" rel=\"nofollow\">migration guide of Runner in MMEngine</a> for more details.</li>\n<li>The Runner in MMEngine also supports testing and validation. The testing scripts are also simplified, which has similar logic to that in training scripts to build the runner.</li>\n<li>The execution points of hooks in the new Runner have been enriched to allow more flexible customization. Please refer to the <a href=\"https://mmengine.readthedocs.io/en/latest/migration/hook.html\" rel=\"nofollow\">migration guide of Hook in MMEngine</a> for more details.</li>\n<li>Learning rate and momentum schedules have been migrated from Hook to <a href=\"https://mmengine.readthedocs.io/en/latest/tutorials/param_scheduler.html\" rel=\"nofollow\">Parameter Scheduler in MMEngine</a>. Please refer to the <a href=\"https://mmengine.readthedocs.io/en/latest/migration/param_scheduler.html\" rel=\"nofollow\">migration guide of Parameter Scheduler in MMEngine</a> for more details.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Configs</h3><a id=\"user-content-configs\" class=\"anchor\" aria-label=\"Permalink: Configs\" href=\"#configs\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>The <a href=\"https://github.com/open-mmlab/mmengine/blob/main/mmengine/runner/runner.py\">Runner in MMEngine</a> uses a different config structure to ease the understanding of the components in the runner. Users can read the <a href=\"../user_guides/config.md\">config example of MMDet 3.x</a> or refer to the <a href=\"https://mmengine.readthedocs.io/en/latest/migration/runner.html\" rel=\"nofollow\">migration guide in MMEngine</a> for migration details.</li>\n<li>The file names of configs and models are also refactored to follow the new rules unified across OpenMMLab 2.0 projects. The names of checkpoints are not updated for now as there is no BC-breaking of model weights between MMDet 3.x and 2.x. We will progressively replace all the model weights with those trained in MMDet 3.x. Please refer to the <a href=\"../user_guides/config.md\">user guides of config</a> for more details.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Dataset</h3><a id=\"user-content-dataset\" class=\"anchor\" aria-label=\"Permalink: Dataset\" href=\"#dataset\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The Dataset classes implemented in MMDet 3.x all inherit from the <code>BaseDetDataset</code>, which inherits from the <a href=\"https://mmengine.readthedocs.io/en/latest/advanced_tutorials/basedataset.html\" rel=\"nofollow\">BaseDataset in MMEngine</a>. In addition to the changes in interfaces, there are several changes in Dataset in MMDet 3.x.</p>\n<ul>\n<li>All the datasets support serializing the internal data list to reduce the memory when multiple workers are built for data loading.</li>\n<li>The internal data structure in the dataset is changed to be self-contained (without losing information like class names in MMDet 2.x) while keeping simplicity.</li>\n<li>The evaluation functionality of each dataset has been removed from the dataset so that some specific evaluation metrics like COCO AP can be used to evaluate the prediction on other datasets.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Data Transforms</h3><a id=\"user-content-data-transforms\" class=\"anchor\" aria-label=\"Permalink: Data Transforms\" href=\"#data-transforms\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The data transforms in MMDet 3.x all inherits from <code>BaseTransform</code> in MMCV&gt;=2.0.0rc0, which defines a new convention in OpenMMLab 2.0 projects.\nBesides the interface changes, there are several changes listed below:</p>\n<ul>\n<li>The functionality of some data transforms (e.g., <code>Resize</code>) are decomposed into several transforms to simplify and clarify the usages.</li>\n<li>The format of data dict processed by each data transform is changed according to the new data structure of dataset.</li>\n<li>Some inefficient data transforms (e.g., normalization and padding) are moved into the data preprocessor of the model to improve data loading and training speed.</li>\n<li>The same data transforms in different OpenMMLab 2.0 libraries have the same augmentation implementation and the logic given the same arguments, i.e., <code>Resize</code> in MMDet 3.x and MMSeg 1.x will resize the image in the exact same manner given the same arguments.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Model</h3><a id=\"user-content-model\" class=\"anchor\" aria-label=\"Permalink: Model\" href=\"#model\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The models in MMDet 3.x all inherit from <code>BaseModel</code> in MMEngine, which defines a new convention of models in OpenMMLab 2.0 projects.\nUsers can refer to <a href=\"https://mmengine.readthedocs.io/en/latest/tutorials/model.html\" rel=\"nofollow\">the tutorial of the model in MMengine</a> for more details.\nAccordingly, there are several changes as the following:</p>\n<ul>\n<li>The model interfaces, including the input and output formats, are significantly simplified and unified following the new convention in MMDet 3.x.\nSpecifically, all the input data in training and testing are packed into <code>inputs</code> and <code>data_samples</code>, where <code>inputs</code> contains model inputs like a list of image tensors, and <code>data_samples</code> contains other information of the current data sample such as ground truths, region proposals, and model predictions. In this way, different tasks in MMDet 3.x can share the same input arguments, which makes the models more general and suitable for multi-task learning and some flexible training paradigms like semi-supervised learning.</li>\n<li>The model has a data preprocessor module, which is used to pre-process the input data of the model. In MMDet 3.x, the data preprocessor usually does the necessary steps to form the input images into a batch, such as padding. It can also serve as a place for some special data augmentations or more efficient data transformations like normalization.</li>\n<li>The internal logic of the model has been changed. In MMdet 2.x, model uses <code>forward_train</code>, <code>forward_test</code>, <code>simple_test</code>, and <code>aug_test</code> to deal with different model forward logics. In MMDet 3.x and OpenMMLab 2.0, the forward function has three modes: 'loss', 'predict', and 'tensor' for training, inference, and tracing or other purposes, respectively.\nThe forward function calls <code>self.loss</code>, <code>self.predict</code>, and <code>self._forward</code> given the modes 'loss', 'predict', and 'tensor', respectively.</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Evaluation</h3><a id=\"user-content-evaluation\" class=\"anchor\" aria-label=\"Permalink: Evaluation\" href=\"#evaluation\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The evaluation in MMDet 2.x strictly binds with the dataset. In contrast, MMDet 3.x decomposes the evaluation from the dataset so that all the detection datasets can evaluate with COCO AP and other metrics implemented in MMDet 3.x.\nMMDet 3.x mainly implements corresponding metrics for each dataset, which are manipulated by <a href=\"https://mmengine.readthedocs.io/en/latest/design/evaluator.html\" rel=\"nofollow\">Evaluator</a> to complete the evaluation.\nUsers can build an evaluator in MMDet 3.x to conduct offline evaluation, i.e., evaluate predictions that may not produce in MMDet 3.x with the dataset as long as the dataset and the prediction follow the dataset conventions. More details can be found in the <a href=\"https://mmengine.readthedocs.io/en/latest/tutorials/evaluation.html\" rel=\"nofollow\">tutorial in mmengine</a>.</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Visualization</h3><a id=\"user-content-visualization\" class=\"anchor\" aria-label=\"Permalink: Visualization\" href=\"#visualization\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>The functions of visualization in MMDet 2.x are removed. Instead, in OpenMMLab 2.0 projects, we use <a href=\"https://mmengine.readthedocs.io/en/latest/design/visualization.html\" rel=\"nofollow\">Visualizer</a> to visualize data. MMDet 3.x implements <code>DetLocalVisualizer</code> to allow visualization of ground truths, model predictions, feature maps, etc., at any place. It also supports sending the visualization data to any external visualization backends such as Tensorboard.</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Optimized training and testing speed of FCOS, RetinaNet, Faster R-CNN, Mask R-CNN, and Cascade R-CNN. The training speed of those models with some common training strategies is also optimized, including those with synchronized batch normalization and mixed precision training.</li>\n<li>Support mixed precision training of all the models. However, some models may get Nan results due to some numerical issues. We will update the documentation and list the results (accuracy of failure) of mixed precision training.</li>\n<li>Release strong baselines of some popular object detectors. Their accuracy and pre-trained checkpoints will be released.</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>DeepFashion dataset: the config and results have been updated.</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ol>\n<li>Support a general semi-supervised learning framework that works with all the object detectors supported in MMDet 3.x. Please refer to <a href=\"../user_guides/semi_det.md\">semi-supervised object detection</a> for details.</li>\n<li>Enable all the single-stage detectors to serve as region proposal networks. We give <a href=\"../user_guides/single_stage_as_rpn.md\">an example of using FCOS as RPN</a>.</li>\n<li>Support a semi-supervised object detection algorithm: <a href=\"https://arxiv.org/abs/2106.09018\" rel=\"nofollow\">SoftTeacher</a>.</li>\n<li>Support <a href=\"https://arxiv.org/abs/2103.07461\" rel=\"nofollow\">the updated CenterNet</a>.</li>\n<li>Support data structures <code>HorizontalBoxes</code> and <code>BaseBoxes</code> to encapsulate different kinds of bounding boxes. We are migrating to use data structures of boxes to replace the use of pure tensor boxes. This will unify the usages of different kinds of bounding boxes in MMDet 3.x and MMRotate 1.x to simplify the implementation and reduce redundant codes.</li>\n</ol>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Planned changes</h2><a id=\"user-content-planned-changes\" class=\"anchor\" aria-label=\"Permalink: Planned changes\" href=\"#planned-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>We list several planned changes of MMDet 3.0.0rc0 so that the community could more comprehensively know the progress of MMDet 3.x. Feel free to create a PR, issue, or discussion if you are interested, have any suggestions and feedback, or want to participate.</p>\n<ol>\n<li>Test-time augmentation: which is supported in MMDet 2.x, is not implemented in this version due to the limited time slot. We will support it in the following releases with a new and simplified design.</li>\n<li>Inference interfaces: unified inference interfaces will be supported in the future to ease the use of released models.</li>\n<li>Interfaces of useful tools that can be used in Jupyter Notebook or Colab: more useful tools that are implemented in the <code>tools</code> directory will have their python interfaces so that they can be used in Jupyter Notebook, Colab, and downstream libraries.</li>\n<li>Documentation: we will add more design docs, tutorials, and migration guidance so that the community can deep dive into our new design, participate the future development, and smoothly migrate downstream libraries to MMDet 3.x.</li>\n<li>Wandb visualization: MMDet 2.x supports data visualization since v2.25.0, which has not been migrated to MMDet 3.x for now. Since WandB provides strong visualization and experiment management capabilities, a <code>DetWandbVisualizer</code> and maybe a hook are planned to fully migrate those functionalities from MMDet 2.x.</li>\n<li>Full support of WiderFace dataset (#8508) and Fast R-CNN: we are verifying their functionalities and will fix related issues soon.</li>\n<li>Migrate DETR-series algorithms (#8655, #8533) and YOLOv3 on IPU (#8552) from MMDet 2.x.</li>\n</ol>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 11 developers contributed to this release.\nThanks @shuxp, @wanghonglie, @Czm369, @BIGWangYuDong, @zytx121, @jbwang1997, @chhluo, @jshilong, @RangiLyu, @hhaAndroid, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@shuxp made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8612\">https://github.com/open-mmlab/mmdetection/pull/8612</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v3.0.0rc0\">https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v3.0.0rc0</a></p>\n","assets":[]},{"id":73079579,"tag":"v2.25.1","author":"ZwwWayne","name":"MMDetection V2.25.1 Release","draft":false,"prerelease":false,"createdAt":"2022-07-28T06:57:27Z","publishedAt":"2022-07-28T15:43:37Z","markdown":"### Bug Fixes\r\n\r\n- Fix single GPU distributed training of cuda device specifying (#8176)\r\n- Fix PolygonMask bug in FilterAnnotations (#8136)\r\n- Fix mdformat version to support python3.6 (#8195)\r\n- Fix GPG key error in Dockerfile (#8215)\r\n- Fix `WandbLoggerHook` error (#8273)\r\n- Fix Pytorch 1.10 incompatibility issues (#8439)\r\n\r\n### Improvements\r\n\r\n- Add `mim` to `extras_require` in setup.py (#8194)\r\n- Support get image shape on macOS (#8434)\r\n- Add test commands of `mim` in CI (#8230 & #8240)\r\n- Update `maskformer` to be compatible when cfg is a dictionary (#8263)\r\n- Clean `Pillow` version check in CI (#8229)\r\n\r\n### Documents\r\n\r\n- Change example hook name in tutorials (#8118)\r\n- Update projects (#8120)\r\n- Update metafile and release new models (#8294)\r\n- Add download link in tutorials (#8391)\r\n\r\n### Contributors\r\n\r\nA total of 15 developers contributed to this release.\r\nThanks @ZwwWayne, @ayulockin, @Mxbonn, @p-mishra1, @Youth-Got, @MiXaiLL76, @chhluo, @jbwang1997, @atinfinity, @shinya7y, @duanzhihua, @STLAND-admin, @BIGWangYuDong, @grimoire, @xiaoyuan0203\r\n\r\n## New Contributors\r\n* @p-mishra1 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8118\r\n* @MiXaiLL76 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8136\r\n* @atinfinity made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8215\r\n* @duanzhihua made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8263\r\n* @STLAND-admin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8370\r\n* @xiaoyuan0203 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/8391\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v2.25.1","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix single GPU distributed training of cuda device specifying (#8176)</li>\n<li>Fix PolygonMask bug in FilterAnnotations (#8136)</li>\n<li>Fix mdformat version to support python3.6 (#8195)</li>\n<li>Fix GPG key error in Dockerfile (#8215)</li>\n<li>Fix <code>WandbLoggerHook</code> error (#8273)</li>\n<li>Fix Pytorch 1.10 incompatibility issues (#8439)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Add <code>mim</code> to <code>extras_require</code> in setup.py (#8194)</li>\n<li>Support get image shape on macOS (#8434)</li>\n<li>Add test commands of <code>mim</code> in CI (#8230 &amp; #8240)</li>\n<li>Update <code>maskformer</code> to be compatible when cfg is a dictionary (#8263)</li>\n<li>Clean <code>Pillow</code> version check in CI (#8229)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Documents</h3><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Change example hook name in tutorials (#8118)</li>\n<li>Update projects (#8120)</li>\n<li>Update metafile and release new models (#8294)</li>\n<li>Add download link in tutorials (#8391)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 15 developers contributed to this release.\nThanks @ZwwWayne, @ayulockin, @Mxbonn, @p-mishra1, @Youth-Got, @MiXaiLL76, @chhluo, @jbwang1997, @atinfinity, @shinya7y, @duanzhihua, @STLAND-admin, @BIGWangYuDong, @grimoire, @xiaoyuan0203</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@p-mishra1 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8118\">https://github.com/open-mmlab/mmdetection/pull/8118</a>\n</li>\n<li>@MiXaiLL76 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8136\">https://github.com/open-mmlab/mmdetection/pull/8136</a>\n</li>\n<li>@atinfinity made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8215\">https://github.com/open-mmlab/mmdetection/pull/8215</a>\n</li>\n<li>@duanzhihua made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8263\">https://github.com/open-mmlab/mmdetection/pull/8263</a>\n</li>\n<li>@STLAND-admin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8370\">https://github.com/open-mmlab/mmdetection/pull/8370</a>\n</li>\n<li>@xiaoyuan0203 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/8391\">https://github.com/open-mmlab/mmdetection/pull/8391</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v2.25.1\">https://github.com/open-mmlab/mmdetection/compare/v2.25.0...v2.25.1</a></p>\n","assets":[]},{"id":68348393,"tag":"v2.25.0","author":"ZwwWayne","name":"MMDetection V2.25.0 Release","draft":false,"prerelease":false,"createdAt":"2022-06-01T14:14:31Z","publishedAt":"2022-06-01T14:16:46Z","markdown":"### Highlights\r\n\r\n- Support dedicated `WandbLogger` hook\r\n- Support [ConvNeXt](configs/convnext), [DDOD](configs/ddod), [SOLOv2](configs/solov2)\r\n- Support [Mask2Former](configs/mask2former) for instance segmentation\r\n- Rename [config files of Mask2Former](configs/mask2former)\r\n\r\n### Backwards incompatible changes\r\n\r\n- Rename [config files of Mask2Former](configs/mask2former) (#7571)\r\n\r\n  <table align=\"center\">\r\n    <thead>\r\n        <tr align='center'>\r\n            <td>before v2.25.0</td>\r\n            <td>after v2.25.0</td>\r\n        </tr>\r\n    </thead>\r\n    <tbody><tr valign='top'>\r\n    <th>\r\n\r\n  - `mask2former_xxx_coco.py` represents config files for **panoptic segmentation**.\r\n\r\n  </th>\r\n    <th>\r\n\r\n  - `mask2former_xxx_coco.py` represents config files for **instance segmentation**.\r\n  - `mask2former_xxx_coco-panoptic.py` represents config files for **panoptic segmentation**.\r\n\r\n  </th></tr>\r\n  </tbody></table>\r\n\r\n### New Features\r\n\r\n- Support [ConvNeXt](https://arxiv.org/abs/2201.03545) (#7281)\r\n- Support [DDOD](https://arxiv.org/abs/2107.02963) (#7279)\r\n- Support [SOLOv2](https://arxiv.org/abs/2003.10152) (#7441)\r\n- Support [Mask2Former](https://arxiv.org/abs/2112.01527) for instance segmentation (#7571, #8032)\r\n\r\n### Bug Fixes\r\n\r\n- Enable YOLOX training on different devices (#7912)\r\n- Fix the log plot error when evaluation with `interval != 1` (#7784)\r\n- Fix RuntimeError of HTC (#8083)\r\n\r\n### Improvements\r\n\r\n- Support dedicated `WandbLogger` hook (#7459)\r\n\r\n  Users can set\r\n\r\n  ```python\r\n  cfg.log_config.hooks = [\r\n    dict(type='MMDetWandbHook',\r\n         init_kwargs={'project': 'MMDetection-tutorial'},\r\n         interval=10,\r\n         log_checkpoint=True,\r\n         log_checkpoint_metadata=True,\r\n         num_eval_images=10)]\r\n  ```\r\n\r\n  in the config to use `MMDetWandbHook`. Example can be found in this [colab tutorial](https://colab.research.google.com/drive/1RCSXHZwDZvakFh3eo9RuNrJbCGqD0dru?usp=sharing#scrollTo=WTEdPDRaBz2C)\r\n\r\n- Add `AvoidOOM` to avoid OOM (#7434, #8091)\r\n\r\n  Try to use `AvoidCUDAOOM` to avoid GPU out of memory. It will first retry after calling `torch.cuda.empty_cache()`. If it still fails, it will then retry by converting the type of inputs to FP16 format. If it still fails, it will try to copy inputs from GPUs to CPUs to continue computing. Try AvoidOOM in code to make the code continue to run when GPU memory runs out:\r\n\r\n  ```python\r\n  from mmdet.utils import AvoidCUDAOOM\r\n\r\n  output = AvoidCUDAOOM.retry_if_cuda_oom(some_function)(input1, input2)\r\n  ```\r\n\r\n  Users can also try `AvoidCUDAOOM` as a decorator to make the code continue to run when GPU memory runs out:\r\n\r\n  ```python\r\n  from mmdet.utils import AvoidCUDAOOM\r\n\r\n  @AvoidCUDAOOM.retry_if_cuda_oom\r\n  def function(*args, **kwargs):\r\n      ...\r\n      return xxx\r\n  ```\r\n\r\n- Support reading `gpu_collect` from `cfg.evaluation.gpu_collect` (#7672)\r\n\r\n- Speedup the Video Inference by Accelerating data-loading Stage (#7832)\r\n\r\n- Support replacing the `${key}` with the value of `cfg.key` (#7492)\r\n\r\n- Accelerate result analysis in `analyze_result.py`. The evaluation time is speedup by 10 ~ 15 times and only tasks 10 ~ 15 minutes now. (#7891)\r\n\r\n- Support to set `block_dilations` in `DilatedEncoder` (#7812)\r\n\r\n- Support panoptic segmentation result analysis (#7922)\r\n\r\n- Release DyHead with Swin-Large backbone (#7733)\r\n\r\n- Documentations updating and adding\r\n\r\n  - Fix wrong default type of `act_cfg` in `SwinTransformer` (#7794)\r\n  - Fix text errors in the tutorials (#7959)\r\n  - Rewrite the [installation guide](docs/en/get_started.md) (#7897)\r\n  - [Useful hooks](docs/en/tutorials/useful_hooks.md) (#7810)\r\n  - Fix heading anchor in documentation  (#8006)\r\n  - Replace `markdownlint` with `mdformat` for avoiding installing ruby (#8009)\r\n\r\n### Contributors\r\n\r\nA total of 20 developers contributed to this release.\r\n\r\nThanks @ZwwWayne, @DarthThomas, @solyaH, @LutingWang, @chenxinfeng4, @Czm369, @Chenastron, @chhluo, @austinmw, @Shanyaliux @hellock, @Y-M-Y, @jbwang1997, @hhaAndroid, @Irvingao, @zhanggefan, @BIGWangYuDong, @Keiku, @PeterVennerstrom, @ayulockin\r\n\r\n## New Contributors\r\n* @DarthThomas made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7794\r\n* @solyaH made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7912\r\n* @chenxinfeng4 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7832\r\n* @Chenastron made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7959\r\n* @austinmw made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7937\r\n* @Shanyaliux made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7812\r\n* @Irvingao made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7279\r\n* @zhanggefan made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7441\r\n* @ayulockin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7459\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.24.1...v2.25.0","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support dedicated <code>WandbLogger</code> hook</li>\n<li>Support <a href=\"configs/convnext\">ConvNeXt</a>, <a href=\"configs/ddod\">DDOD</a>, <a href=\"configs/solov2\">SOLOv2</a>\n</li>\n<li>Support <a href=\"configs/mask2former\">Mask2Former</a> for instance segmentation</li>\n<li>Rename <a href=\"configs/mask2former\">config files of Mask2Former</a>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Backwards incompatible changes</h3><a id=\"user-content-backwards-incompatible-changes\" class=\"anchor\" aria-label=\"Permalink: Backwards incompatible changes\" href=\"#backwards-incompatible-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>\n<p>Rename <a href=\"configs/mask2former\">config files of Mask2Former</a> (#7571)</p>\n<table align=\"center\">\n  <thead>\n      <tr align=\"center\">\n          <td>before v2.25.0</td>\n          <td>after v2.25.0</td>\n      </tr>\n  </thead>\n  <tbody>\n<tr valign=\"top\">\n  <th>\n<ul>\n<li>\n<code>mask2former_xxx_coco.py</code> represents config files for <strong>panoptic segmentation</strong>.</li>\n</ul>\n</th>\n  <th>\n<ul>\n<li>\n<code>mask2former_xxx_coco.py</code> represents config files for <strong>instance segmentation</strong>.</li>\n<li>\n<code>mask2former_xxx_coco-panoptic.py</code> represents config files for <strong>panoptic segmentation</strong>.</li>\n</ul>\n</th>\n</tr>\n</tbody>\n</table>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2201.03545\" rel=\"nofollow\">ConvNeXt</a> (#7281)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2107.02963\" rel=\"nofollow\">DDOD</a> (#7279)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2003.10152\" rel=\"nofollow\">SOLOv2</a> (#7441)</li>\n<li>Support <a href=\"https://arxiv.org/abs/2112.01527\" rel=\"nofollow\">Mask2Former</a> for instance segmentation (#7571, #8032)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Enable YOLOX training on different devices (#7912)</li>\n<li>Fix the log plot error when evaluation with <code>interval != 1</code> (#7784)</li>\n<li>Fix RuntimeError of HTC (#8083)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>\n<p>Support dedicated <code>WandbLogger</code> hook (#7459)</p>\n<p>Users can set</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-s1\">cfg</span>.<span class=\"pl-c1\">log_config</span>.<span class=\"pl-c1\">hooks</span> <span class=\"pl-c1\">=</span> [\n  <span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'MMDetWandbHook'</span>,\n       <span class=\"pl-s1\">init_kwargs</span><span class=\"pl-c1\">=</span>{<span class=\"pl-s\">'project'</span>: <span class=\"pl-s\">'MMDetection-tutorial'</span>},\n       <span class=\"pl-s1\">interval</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">10</span>,\n       <span class=\"pl-s1\">log_checkpoint</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">True</span>,\n       <span class=\"pl-s1\">log_checkpoint_metadata</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">True</span>,\n       <span class=\"pl-s1\">num_eval_images</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">10</span>)]</pre></div>\n<p>in the config to use <code>MMDetWandbHook</code>. Example can be found in this <a href=\"https://colab.research.google.com/drive/1RCSXHZwDZvakFh3eo9RuNrJbCGqD0dru?usp=sharing#scrollTo=WTEdPDRaBz2C\" rel=\"nofollow\">colab tutorial</a></p>\n</li>\n<li>\n<p>Add <code>AvoidOOM</code> to avoid OOM (#7434, #8091)</p>\n<p>Try to use <code>AvoidCUDAOOM</code> to avoid GPU out of memory. It will first retry after calling <code>torch.cuda.empty_cache()</code>. If it still fails, it will then retry by converting the type of inputs to FP16 format. If it still fails, it will try to copy inputs from GPUs to CPUs to continue computing. Try AvoidOOM in code to make the code continue to run when GPU memory runs out:</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-k\">from</span> <span class=\"pl-s1\">mmdet</span>.<span class=\"pl-s1\">utils</span> <span class=\"pl-k\">import</span> <span class=\"pl-v\">AvoidCUDAOOM</span>\n\n<span class=\"pl-s1\">output</span> <span class=\"pl-c1\">=</span> <span class=\"pl-v\">AvoidCUDAOOM</span>.<span class=\"pl-c1\">retry_if_cuda_oom</span>(<span class=\"pl-s1\">some_function</span>)(<span class=\"pl-s1\">input1</span>, <span class=\"pl-s1\">input2</span>)</pre></div>\n<p>Users can also try <code>AvoidCUDAOOM</code> as a decorator to make the code continue to run when GPU memory runs out:</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-k\">from</span> <span class=\"pl-s1\">mmdet</span>.<span class=\"pl-s1\">utils</span> <span class=\"pl-k\">import</span> <span class=\"pl-v\">AvoidCUDAOOM</span>\n\n<span class=\"pl-en\">@<span class=\"pl-v\">AvoidCUDAOOM</span>.<span class=\"pl-c1\">retry_if_cuda_oom</span></span>\n<span class=\"pl-k\">def</span> <span class=\"pl-en\">function</span>(<span class=\"pl-c1\">*</span><span class=\"pl-s1\">args</span>, <span class=\"pl-c1\">**</span><span class=\"pl-s1\">kwargs</span>):\n    ...\n    <span class=\"pl-k\">return</span> <span class=\"pl-s1\">xxx</span></pre></div>\n</li>\n<li>\n<p>Support reading <code>gpu_collect</code> from <code>cfg.evaluation.gpu_collect</code> (#7672)</p>\n</li>\n<li>\n<p>Speedup the Video Inference by Accelerating data-loading Stage (#7832)</p>\n</li>\n<li>\n<p>Support replacing the <code>${key}</code> with the value of <code>cfg.key</code> (#7492)</p>\n</li>\n<li>\n<p>Accelerate result analysis in <code>analyze_result.py</code>. The evaluation time is speedup by 10 ~ 15 times and only tasks 10 ~ 15 minutes now. (#7891)</p>\n</li>\n<li>\n<p>Support to set <code>block_dilations</code> in <code>DilatedEncoder</code> (#7812)</p>\n</li>\n<li>\n<p>Support panoptic segmentation result analysis (#7922)</p>\n</li>\n<li>\n<p>Release DyHead with Swin-Large backbone (#7733)</p>\n</li>\n<li>\n<p>Documentations updating and adding</p>\n<ul>\n<li>Fix wrong default type of <code>act_cfg</code> in <code>SwinTransformer</code> (#7794)</li>\n<li>Fix text errors in the tutorials (#7959)</li>\n<li>Rewrite the <a href=\"docs/en/get_started.md\">installation guide</a> (#7897)</li>\n<li>\n<a href=\"docs/en/tutorials/useful_hooks.md\">Useful hooks</a> (#7810)</li>\n<li>Fix heading anchor in documentation  (#8006)</li>\n<li>Replace <code>markdownlint</code> with <code>mdformat</code> for avoiding installing ruby (#8009)</li>\n</ul>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 20 developers contributed to this release.</p>\n<p>Thanks @ZwwWayne, @DarthThomas, @solyaH, @LutingWang, @chenxinfeng4, @Czm369, @Chenastron, @chhluo, @austinmw, @Shanyaliux @hellock, @Y-M-Y, @jbwang1997, @hhaAndroid, @Irvingao, @zhanggefan, @BIGWangYuDong, @Keiku, @PeterVennerstrom, @ayulockin</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@DarthThomas made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7794\">https://github.com/open-mmlab/mmdetection/pull/7794</a>\n</li>\n<li>@solyaH made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7912\">https://github.com/open-mmlab/mmdetection/pull/7912</a>\n</li>\n<li>@chenxinfeng4 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7832\">https://github.com/open-mmlab/mmdetection/pull/7832</a>\n</li>\n<li>@Chenastron made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7959\">https://github.com/open-mmlab/mmdetection/pull/7959</a>\n</li>\n<li>@austinmw made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7937\">https://github.com/open-mmlab/mmdetection/pull/7937</a>\n</li>\n<li>@Shanyaliux made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7812\">https://github.com/open-mmlab/mmdetection/pull/7812</a>\n</li>\n<li>@Irvingao made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7279\">https://github.com/open-mmlab/mmdetection/pull/7279</a>\n</li>\n<li>@zhanggefan made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7441\">https://github.com/open-mmlab/mmdetection/pull/7441</a>\n</li>\n<li>@ayulockin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7459\">https://github.com/open-mmlab/mmdetection/pull/7459</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.24.1...v2.25.0\">https://github.com/open-mmlab/mmdetection/compare/v2.24.1...v2.25.0</a></p>\n","assets":[]},{"id":65681744,"tag":"v2.24.1","author":"ZwwWayne","name":"MMDetection V2.24.1 Release","draft":false,"prerelease":false,"createdAt":"2022-04-30T14:23:14Z","publishedAt":"2022-04-30T14:27:27Z","markdown":"## What's Changed\r\n* [Fix] Fix no attribute 'samples_per_gpu' bug in `auto_scale_lr` by @jbwang1997 in https://github.com/open-mmlab/mmdetection/pull/7862\r\n* [Fix] Fix KeyError: 'ori_filename' when using --show-dir with centernet_resnet18_dcnv2_140e_coco.py by @jbwang1997 in https://github.com/open-mmlab/mmdetection/pull/7865\r\n* [Fix] Fix the configs of simplecopypaste by @Czm369 in https://github.com/open-mmlab/mmdetection/pull/7864\r\n* [Docs] Update readme by @chhluo in https://github.com/open-mmlab/mmdetection/pull/7867\r\n\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.24.0...v2.24.1","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">What's Changed</h2><a id=\"user-content-whats-changed\" class=\"anchor\" aria-label=\"Permalink: What's Changed\" href=\"#whats-changed\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>[Fix] Fix no attribute 'samples_per_gpu' bug in <code>auto_scale_lr</code> by @jbwang1997 in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7862\">https://github.com/open-mmlab/mmdetection/pull/7862</a>\n</li>\n<li>[Fix] Fix KeyError: 'ori_filename' when using --show-dir with centernet_resnet18_dcnv2_140e_coco.py by @jbwang1997 in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7865\">https://github.com/open-mmlab/mmdetection/pull/7865</a>\n</li>\n<li>[Fix] Fix the configs of simplecopypaste by @Czm369 in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7864\">https://github.com/open-mmlab/mmdetection/pull/7864</a>\n</li>\n<li>[Docs] Update readme by @chhluo in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7867\">https://github.com/open-mmlab/mmdetection/pull/7867</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.24.0...v2.24.1\">https://github.com/open-mmlab/mmdetection/compare/v2.24.0...v2.24.1</a></p>\n","assets":[]},{"id":65355269,"tag":"v2.24.0","author":"ZwwWayne","name":"MMDetection V2.24.0 Release","draft":false,"prerelease":false,"createdAt":"2022-04-26T13:14:55Z","publishedAt":"2022-04-26T14:08:40Z","markdown":"#### Highlights\r\n\r\n- Support [Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation](https://arxiv.org/abs/2012.07177)\r\n- Support automatically scaling LR according to GPU number and samples per GPU\r\n- Support Class Aware Sampler that improves performance on OpenImages Dataset\r\n\r\n#### New Features\r\n\r\n- Support [Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation](https://arxiv.org/abs/2012.07177), see [example configs](configs/simple_copy_paste/mask_rcnn_r50_fpn_syncbn-all_rpn-2conv_ssj_scp_32x2_270k_coco.py) (#7501)\r\n- Support Class Aware Sampler, users can set\r\n\r\n  ```python\r\n  data=dict(train_dataloader=dict(class_aware_sampler=dict(num_sample_class=1))))\r\n  ```\r\n\r\n  in the config to use `ClassAwareSampler`. Examples can be found in [the configs of OpenImages Dataset](https://github.com/open-mmlab/mmdetection/tree/master/configs/openimages/faster_rcnn_r50_fpn_32x2_cas_1x_openimages.py).  (#7436)\r\n\r\n- Support automatically scaling LR according to GPU number and samples per GPU. (#7482)\r\n  In each config, there is a corresponding config of auto-scaling LR as below,\r\n\r\n   ```python\r\n   auto_scale_lr = dict(enable=True, base_batch_size=N)\r\n   ```\r\n\r\n  where `N` is the batch size used for the current learning rate in the config (also equals to `samples_per_gpu` * gpu number to train this config).\r\n  By default, we set `enable=False` so that the original usages will not be affected. Users can set `enable=True` in each config or add `--auto-scale-lr` after the command line to enable this feature and should check the correctness of `base_batch_size` in customized configs.\r\n\r\n- Support setting dataloader arguments in config and add functions to handle config compatibility. (#7668)\r\n  The comparison between the old and new usages is as below.\r\n\r\n  <table align=\"center\">\r\n    <thead>\r\n        <tr align='center'>\r\n            <td>Before v2.24.0</td>\r\n            <td>Since v2.24.0 </td>\r\n        </tr>\r\n    </thead>\r\n    <tbody><tr valign='top'>\r\n    <th>\r\n\r\n    ```python\r\n    data = dict(\r\n        samples_per_gpu=64, workers_per_gpu=4,\r\n        train=dict(type='xxx', ...),\r\n        val=dict(type='xxx', samples_per_gpu=4, ...),\r\n        test=dict(type='xxx', ...),\r\n    )\r\n    ```\r\n\r\n    </th>\r\n    <th>\r\n\r\n    ```python\r\n    # A recommended config that is clear\r\n    data = dict(\r\n        train=dict(type='xxx', ...),\r\n        val=dict(type='xxx', ...),\r\n        test=dict(type='xxx', ...),\r\n        # Use different batch size during inference.\r\n        train_dataloader=dict(samples_per_gpu=64, workers_per_gpu=4),\r\n        val_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2),\r\n        test_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2),\r\n    )\r\n\r\n    # Old style still works but allows to set more arguments about data loaders\r\n    data = dict(\r\n        samples_per_gpu=64,  # only works for train_dataloader\r\n        workers_per_gpu=4,  # only works for train_dataloader\r\n        train=dict(type='xxx', ...),\r\n        val=dict(type='xxx', ...),\r\n        test=dict(type='xxx', ...),\r\n        # Use different batch size during inference.\r\n        val_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2),\r\n        test_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2),\r\n    )\r\n    ```\r\n\r\n    </th></tr>\r\n  </tbody></table>\r\n\r\n- Support memory profile hook. Users can use it to monitor the memory usages during training as below (#7560)\r\n\r\n    ```python\r\n    custom_hooks = [\r\n        dict(type='MemoryProfilerHook', interval=50)\r\n    ]\r\n    ```\r\n\r\n- Support to run on PyTorch with MLU chip (#7578)\r\n- Support re-spliting data batch with tag (#7641)\r\n- Support the `DiceCost` used by [K-Net](https://arxiv.org/abs/2106.14855) in `MaskHungarianAssigner` (#7716)\r\n- Support splitting COCO data for Semi-supervised object detection (#7431)\r\n- Support Pathlib for Config.fromfile (#7685)\r\n- Support to use file client in OpenImages dataset (#7433)\r\n- Add a probability parameter to Mosaic transformation (#7371)\r\n- Support specifying interpolation mode in `Resize` pipeline (#7585)\r\n\r\n#### Bug Fixes\r\n\r\n- Avoid invalid bbox after deform_sampling (#7567)\r\n- Fix the issue that argument color_theme does not take effect when exporting confusion matrix (#7701)\r\n- Fix the `end_level` in Necks, which should be the index of the end input backbone level (#7502)\r\n- Fix the bug that `mix_results` may be None in `MultiImageMixDataset` (#7530)\r\n- Fix the bug in ResNet plugin when two plugins are used (#7797)\r\n\r\n#### Improvements\r\n\r\n- Enhance `load_json_logs` of analyze_logs.py for resumed training logs (#7732)\r\n- Add argument `out_file` in image_demo.py (#7676)\r\n- Allow mixed precision training with `SimOTAAssigner` (#7516)\r\n- Updated INF to 100000.0 to be the same as that in the official YOLOX (#7778)\r\n- Add documentations of:\r\n  - how to get channels of a new backbone (#7642)\r\n  - how to unfreeze the backbone network (#7570)\r\n  - how to train fast_rcnn model (#7549)\r\n  - proposals in Deformable DETR (#7690)\r\n  - from-scratch install script in get_started.md (#7575)\r\n- Release pre-trained models of\r\n  - [Mask2Former](configs/mask2former) (#7595, #7709)\r\n  - RetinaNet with ResNet-18 and release models (#7387)\r\n  - RetinaNet with EfficientNet backbone (#7646)\r\n\r\n#### Contributors\r\n\r\nA total of 27 developers contributed to this release.\r\nThanks @jovialio, @zhangsanfeng2022, @HarryZJ, @jamiechoi1995, @nestiank, @PeterH0323, @RangeKing, @Y-M-Y, @mattcasey02, @weiji14, @Yulv-git, @xiefeifeihu, @FANG-MING, @meng976537406, @nijkah, @sudz123, @CCODING04, @SheffieldCao, @Czm369, @BIGWangYuDong, @zytx121, @jbwang1997, @chhluo, @jshilong, @RangiLyu, @hhaAndroid, @ZwwWayne\r\n\r\n## New Contributors\r\n* @nestiank made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7591\r\n* @PeterH0323 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7482\r\n* @mattcasey02 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7610\r\n* @weiji14 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7516\r\n* @Yulv-git made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7679\r\n* @xiefeifeihu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7701\r\n* @SheffieldCao made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7732\r\n* @jovialio made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7778\r\n* @zhangsanfeng2022 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7578\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.23.0...v2.24.0","html":"<div class=\"markdown-heading\"><h4 class=\"heading-element\">Highlights</h4><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2012.07177\" rel=\"nofollow\">Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation</a>\n</li>\n<li>Support automatically scaling LR according to GPU number and samples per GPU</li>\n<li>Support Class Aware Sampler that improves performance on OpenImages Dataset</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">New Features</h4><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>\n<p>Support <a href=\"https://arxiv.org/abs/2012.07177\" rel=\"nofollow\">Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation</a>, see <a href=\"configs/simple_copy_paste/mask_rcnn_r50_fpn_syncbn-all_rpn-2conv_ssj_scp_32x2_270k_coco.py\">example configs</a> (#7501)</p>\n</li>\n<li>\n<p>Support Class Aware Sampler, users can set</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-s1\">data</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">train_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">class_aware_sampler</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">num_sample_class</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">1</span>))))</pre></div>\n<p>in the config to use <code>ClassAwareSampler</code>. Examples can be found in <a href=\"https://github.com/open-mmlab/mmdetection/tree/master/configs/openimages/faster_rcnn_r50_fpn_32x2_cas_1x_openimages.py\">the configs of OpenImages Dataset</a>.  (#7436)</p>\n</li>\n<li>\n<p>Support automatically scaling LR according to GPU number and samples per GPU. (#7482)\nIn each config, there is a corresponding config of auto-scaling LR as below,</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-s1\">auto_scale_lr</span> <span class=\"pl-c1\">=</span> <span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">enable</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">True</span>, <span class=\"pl-s1\">base_batch_size</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">N</span>)</pre></div>\n<p>where <code>N</code> is the batch size used for the current learning rate in the config (also equals to <code>samples_per_gpu</code> * gpu number to train this config).\nBy default, we set <code>enable=False</code> so that the original usages will not be affected. Users can set <code>enable=True</code> in each config or add <code>--auto-scale-lr</code> after the command line to enable this feature and should check the correctness of <code>base_batch_size</code> in customized configs.</p>\n</li>\n<li>\n<p>Support setting dataloader arguments in config and add functions to handle config compatibility. (#7668)\nThe comparison between the old and new usages is as below.</p>\n<table align=\"center\">\n  <thead>\n      <tr align=\"center\">\n          <td>Before v2.24.0</td>\n          <td>Since v2.24.0 </td>\n      </tr>\n  </thead>\n  <tbody>\n<tr valign=\"top\">\n  <th>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-s1\">data</span> <span class=\"pl-c1\">=</span> <span class=\"pl-en\">dict</span>(\n    <span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">64</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">4</span>,\n    <span class=\"pl-s1\">train</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-s1\">val</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, <span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">4</span>, ...),\n    <span class=\"pl-s1\">test</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n)</pre></div>\n  </th>\n  <th>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-c\"># A recommended config that is clear</span>\n<span class=\"pl-s1\">data</span> <span class=\"pl-c1\">=</span> <span class=\"pl-en\">dict</span>(\n    <span class=\"pl-s1\">train</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-s1\">val</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-s1\">test</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-c\"># Use different batch size during inference.</span>\n    <span class=\"pl-s1\">train_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">64</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">4</span>),\n    <span class=\"pl-s1\">val_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">8</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">2</span>),\n    <span class=\"pl-s1\">test_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">8</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">2</span>),\n)\n\n<span class=\"pl-c\"># Old style still works but allows to set more arguments about data loaders</span>\n<span class=\"pl-s1\">data</span> <span class=\"pl-c1\">=</span> <span class=\"pl-en\">dict</span>(\n    <span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">64</span>,  <span class=\"pl-c\"># only works for train_dataloader</span>\n    <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">4</span>,  <span class=\"pl-c\"># only works for train_dataloader</span>\n    <span class=\"pl-s1\">train</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-s1\">val</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-s1\">test</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'xxx'</span>, ...),\n    <span class=\"pl-c\"># Use different batch size during inference.</span>\n    <span class=\"pl-s1\">val_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">8</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">2</span>),\n    <span class=\"pl-s1\">test_dataloader</span><span class=\"pl-c1\">=</span><span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">samples_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">8</span>, <span class=\"pl-s1\">workers_per_gpu</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">2</span>),\n)</pre></div>\n  </th>\n</tr>\n</tbody>\n</table>\n</li>\n<li>\n<p>Support memory profile hook. Users can use it to monitor the memory usages during training as below (#7560)</p>\n<div class=\"highlight highlight-source-python\"><pre><span class=\"pl-s1\">custom_hooks</span> <span class=\"pl-c1\">=</span> [\n    <span class=\"pl-en\">dict</span>(<span class=\"pl-s1\">type</span><span class=\"pl-c1\">=</span><span class=\"pl-s\">'MemoryProfilerHook'</span>, <span class=\"pl-s1\">interval</span><span class=\"pl-c1\">=</span><span class=\"pl-c1\">50</span>)\n]</pre></div>\n</li>\n<li>\n<p>Support to run on PyTorch with MLU chip (#7578)</p>\n</li>\n<li>\n<p>Support re-spliting data batch with tag (#7641)</p>\n</li>\n<li>\n<p>Support the <code>DiceCost</code> used by <a href=\"https://arxiv.org/abs/2106.14855\" rel=\"nofollow\">K-Net</a> in <code>MaskHungarianAssigner</code> (#7716)</p>\n</li>\n<li>\n<p>Support splitting COCO data for Semi-supervised object detection (#7431)</p>\n</li>\n<li>\n<p>Support Pathlib for Config.fromfile (#7685)</p>\n</li>\n<li>\n<p>Support to use file client in OpenImages dataset (#7433)</p>\n</li>\n<li>\n<p>Add a probability parameter to Mosaic transformation (#7371)</p>\n</li>\n<li>\n<p>Support specifying interpolation mode in <code>Resize</code> pipeline (#7585)</p>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Bug Fixes</h4><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Avoid invalid bbox after deform_sampling (#7567)</li>\n<li>Fix the issue that argument color_theme does not take effect when exporting confusion matrix (#7701)</li>\n<li>Fix the <code>end_level</code> in Necks, which should be the index of the end input backbone level (#7502)</li>\n<li>Fix the bug that <code>mix_results</code> may be None in <code>MultiImageMixDataset</code> (#7530)</li>\n<li>Fix the bug in ResNet plugin when two plugins are used (#7797)</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Improvements</h4><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Enhance <code>load_json_logs</code> of analyze_logs.py for resumed training logs (#7732)</li>\n<li>Add argument <code>out_file</code> in image_demo.py (#7676)</li>\n<li>Allow mixed precision training with <code>SimOTAAssigner</code> (#7516)</li>\n<li>Updated INF to 100000.0 to be the same as that in the official YOLOX (#7778)</li>\n<li>Add documentations of:\n<ul>\n<li>how to get channels of a new backbone (#7642)</li>\n<li>how to unfreeze the backbone network (#7570)</li>\n<li>how to train fast_rcnn model (#7549)</li>\n<li>proposals in Deformable DETR (#7690)</li>\n<li>from-scratch install script in get_started.md (#7575)</li>\n</ul>\n</li>\n<li>Release pre-trained models of\n<ul>\n<li>\n<a href=\"configs/mask2former\">Mask2Former</a> (#7595, #7709)</li>\n<li>RetinaNet with ResNet-18 and release models (#7387)</li>\n<li>RetinaNet with EfficientNet backbone (#7646)</li>\n</ul>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Contributors</h4><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 27 developers contributed to this release.\nThanks @jovialio, @zhangsanfeng2022, @HarryZJ, @jamiechoi1995, @nestiank, @PeterH0323, @RangeKing, @Y-M-Y, @mattcasey02, @weiji14, @Yulv-git, @xiefeifeihu, @FANG-MING, @meng976537406, @nijkah, @sudz123, @CCODING04, @SheffieldCao, @Czm369, @BIGWangYuDong, @zytx121, @jbwang1997, @chhluo, @jshilong, @RangiLyu, @hhaAndroid, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@nestiank made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7591\">https://github.com/open-mmlab/mmdetection/pull/7591</a>\n</li>\n<li>@PeterH0323 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7482\">https://github.com/open-mmlab/mmdetection/pull/7482</a>\n</li>\n<li>@mattcasey02 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7610\">https://github.com/open-mmlab/mmdetection/pull/7610</a>\n</li>\n<li>@weiji14 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7516\">https://github.com/open-mmlab/mmdetection/pull/7516</a>\n</li>\n<li>@Yulv-git made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7679\">https://github.com/open-mmlab/mmdetection/pull/7679</a>\n</li>\n<li>@xiefeifeihu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7701\">https://github.com/open-mmlab/mmdetection/pull/7701</a>\n</li>\n<li>@SheffieldCao made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7732\">https://github.com/open-mmlab/mmdetection/pull/7732</a>\n</li>\n<li>@jovialio made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7778\">https://github.com/open-mmlab/mmdetection/pull/7778</a>\n</li>\n<li>@zhangsanfeng2022 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7578\">https://github.com/open-mmlab/mmdetection/pull/7578</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.23.0...v2.24.0\">https://github.com/open-mmlab/mmdetection/compare/v2.23.0...v2.24.0</a></p>\n","assets":[]},{"id":63097917,"tag":"v2.23.0","author":"ZwwWayne","name":"MMDetection V2.23.0 Release","draft":false,"prerelease":false,"createdAt":"2022-03-30T01:42:05Z","publishedAt":"2022-03-30T01:42:37Z","markdown":"## Highlights\r\n\r\n- Support Mask2Former: [Masked-attention Mask Transformer for Universal Image Segmentation](https://arxiv.org/abs/2112.01527)\r\n- Support EfficientNet: [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946)\r\n- Support setting data root through environment variable `MMDET_DATASETS`, users don't have to modify the corresponding path in config files anymore.\r\n- Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision.\r\n\r\n## New Features\r\n\r\n- Support [Mask2Former](configs/mask2former)(#6938)(#7466)(#7471)\r\n- Support [EfficientNet](configs/efficientnet) (#7514)\r\n- Support setting data root through environment variable `MMDET_DATASETS`, users don't have to modify the corresponding path in config files anymore. (#7386)\r\n- Support setting different seeds to different ranks (#7432)\r\n- Update the `dist_train.sh` so that the script can be used to support launching multi-node training on machines without slurm (#7415)\r\n- Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision (#7489)\r\n\r\n## Bug Fixes\r\n\r\n- Fix bug in VOC unit test which removes the data directory (#7270)\r\n- Adjust the order of `get_classes` and `FileClient` (#7276)\r\n- Force the inputs of `get_bboxes` in yolox_head to float32 (#7324)\r\n- Fix misplaced arguments in LoadPanopticAnnotations (#7388)\r\n- Fix reduction=mean in CELoss. (#7449)\r\n- Update unit test of CrossEntropyCost (#7537)\r\n- Fix memory leaking in panpotic segmentation evaluation (#7538)\r\n- Fix the bug of shape broadcast in YOLOv3 (#7551)\r\n\r\n## Improvements\r\n\r\n- Add Chinese version of onnx2tensorrt.md (#7219)\r\n- Update colab tutorials (#7310)\r\n- Update information about Localization Distillation (#7350)\r\n- Add Chinese version of `finetune.md` (#7178)\r\n- Update YOLOX log for non square input (#7235)\r\n- Add `nproc` in `coco_panoptic.py` for panoptic quality computing (#7315)\r\n- Allow to set channel_order in LoadImageFromFile (#7258)\r\n- Take point sample related functions out of mask_point_head (#7353)\r\n- Add instance evaluation for coco_panoptic (#7313)\r\n- Enhance the robustness of analyze_logs.py (#7407)\r\n- Supplementary notes of sync_random_seed (#7440)\r\n- Update docstring of cross entropy loss (#7472)\r\n- Update pascal voc result (#7503)\r\n- We create How-to documentation to record any questions about How to xxx. In this version, we added\r\n  - How to use Mosaic augmentation (#7507)\r\n  - How to use backbone in mmcls (#7438)\r\n  - How to produce and submit the prediction results of panoptic segmentation models on COCO test-dev set (#7430))\r\n\r\n## Contributors\r\n\r\nA total of 27 developers contributed to this release.\r\nThanks @ZwwWayne, @haofanwang, @shinya7y, @chhluo, @yangrisheng, @triple-Mu, @jbwang1997, @HikariTJU, @imflash217, @274869388, @zytx121, @matrixgame2018, @jamiechoi1995, @BIGWangYuDong, @JingweiZhang12, @Xiangxu-0103, @hhaAndroid, @jshilong, @osbm, @ceroytres, @bunge-bedstraw-herb, @Youth-Got, @daavoo, @jiangyitong, @RangiLyu, @CCODING04, @yarkable\r\n\r\n\r\n## New Contributors\r\n* @triple-Mu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7219\r\n* @yangrisheng made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7180\r\n* @imflash217 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7398\r\n* @osbm made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7498\r\n* @ceroytres made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7510\r\n* @bunge-bedstraw-herb made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7507\r\n* @Youth-Got made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7521\r\n* @CCODING04 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7386\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.22.0...v2.23.0","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support Mask2Former: <a href=\"https://arxiv.org/abs/2112.01527\" rel=\"nofollow\">Masked-attention Mask Transformer for Universal Image Segmentation</a>\n</li>\n<li>Support EfficientNet: <a href=\"https://arxiv.org/abs/1905.11946\" rel=\"nofollow\">EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks</a>\n</li>\n<li>Support setting data root through environment variable <code>MMDET_DATASETS</code>, users don't have to modify the corresponding path in config files anymore.</li>\n<li>Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision.</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"configs/mask2former\">Mask2Former</a>(#6938)(#7466)(#7471)</li>\n<li>Support <a href=\"configs/efficientnet\">EfficientNet</a> (#7514)</li>\n<li>Support setting data root through environment variable <code>MMDET_DATASETS</code>, users don't have to modify the corresponding path in config files anymore. (#7386)</li>\n<li>Support setting different seeds to different ranks (#7432)</li>\n<li>Update the <code>dist_train.sh</code> so that the script can be used to support launching multi-node training on machines without slurm (#7415)</li>\n<li>Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision (#7489)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix bug in VOC unit test which removes the data directory (#7270)</li>\n<li>Adjust the order of <code>get_classes</code> and <code>FileClient</code> (#7276)</li>\n<li>Force the inputs of <code>get_bboxes</code> in yolox_head to float32 (#7324)</li>\n<li>Fix misplaced arguments in LoadPanopticAnnotations (#7388)</li>\n<li>Fix reduction=mean in CELoss. (#7449)</li>\n<li>Update unit test of CrossEntropyCost (#7537)</li>\n<li>Fix memory leaking in panpotic segmentation evaluation (#7538)</li>\n<li>Fix the bug of shape broadcast in YOLOv3 (#7551)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Add Chinese version of onnx2tensorrt.md (#7219)</li>\n<li>Update colab tutorials (#7310)</li>\n<li>Update information about Localization Distillation (#7350)</li>\n<li>Add Chinese version of <code>finetune.md</code> (#7178)</li>\n<li>Update YOLOX log for non square input (#7235)</li>\n<li>Add <code>nproc</code> in <code>coco_panoptic.py</code> for panoptic quality computing (#7315)</li>\n<li>Allow to set channel_order in LoadImageFromFile (#7258)</li>\n<li>Take point sample related functions out of mask_point_head (#7353)</li>\n<li>Add instance evaluation for coco_panoptic (#7313)</li>\n<li>Enhance the robustness of analyze_logs.py (#7407)</li>\n<li>Supplementary notes of sync_random_seed (#7440)</li>\n<li>Update docstring of cross entropy loss (#7472)</li>\n<li>Update pascal voc result (#7503)</li>\n<li>We create How-to documentation to record any questions about How to xxx. In this version, we added\n<ul>\n<li>How to use Mosaic augmentation (#7507)</li>\n<li>How to use backbone in mmcls (#7438)</li>\n<li>How to produce and submit the prediction results of panoptic segmentation models on COCO test-dev set (#7430))</li>\n</ul>\n</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 27 developers contributed to this release.\nThanks @ZwwWayne, @haofanwang, @shinya7y, @chhluo, @yangrisheng, @triple-Mu, @jbwang1997, @HikariTJU, @imflash217, @274869388, @zytx121, @matrixgame2018, @jamiechoi1995, @BIGWangYuDong, @JingweiZhang12, @Xiangxu-0103, @hhaAndroid, @jshilong, @osbm, @ceroytres, @bunge-bedstraw-herb, @Youth-Got, @daavoo, @jiangyitong, @RangiLyu, @CCODING04, @yarkable</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@triple-Mu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7219\">https://github.com/open-mmlab/mmdetection/pull/7219</a>\n</li>\n<li>@yangrisheng made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7180\">https://github.com/open-mmlab/mmdetection/pull/7180</a>\n</li>\n<li>@imflash217 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7398\">https://github.com/open-mmlab/mmdetection/pull/7398</a>\n</li>\n<li>@osbm made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7498\">https://github.com/open-mmlab/mmdetection/pull/7498</a>\n</li>\n<li>@ceroytres made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7510\">https://github.com/open-mmlab/mmdetection/pull/7510</a>\n</li>\n<li>@bunge-bedstraw-herb made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7507\">https://github.com/open-mmlab/mmdetection/pull/7507</a>\n</li>\n<li>@Youth-Got made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7521\">https://github.com/open-mmlab/mmdetection/pull/7521</a>\n</li>\n<li>@CCODING04 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7386\">https://github.com/open-mmlab/mmdetection/pull/7386</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.22.0...v2.23.0\">https://github.com/open-mmlab/mmdetection/compare/v2.22.0...v2.23.0</a></p>\n","assets":[]},{"id":60525727,"tag":"v2.22.0","author":"ZwwWayne","name":"MMDetection V2.22.0 Release","draft":false,"prerelease":false,"createdAt":"2022-02-26T02:26:32Z","publishedAt":"2022-02-26T02:28:30Z","markdown":"### Breaking Changes\r\n\r\nIn order to support the visualization for Panoptic Segmentation, the `num_classes` can not be `None` when using the `get_palette` function to determine whether to use the panoptic palette.\r\n\r\n### Highlights\r\n\r\n- Support MaskFormer: [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) (#7212)\r\n- Support DyHead: [Dynamic Head: Unifying Object Detection Heads with Attentions](https://arxiv.org/abs/2106.08322) (#6823)\r\n- Release a good recipe of using ResNet in object detectors pre-trained by [ResNet Strikes Back](https://arxiv.org/abs/2110.00476), which consistently brings about 3~4 mAP improvements over RetinaNet, Faster/Mask/Cascade Mask R-CNN (#7001)\r\n- Support [Open Images Dataset](https://storage.googleapis.com/openimages/web/index.html) (#6331)\r\n- Support TIMM backbone: [PyTorch Image Models](https://github.com/rwightman/pytorch-image-models) (#7020)\r\n\r\n### New Features\r\n\r\n- Support [MaskFormer](configs/maskformer) (#7212)\r\n- Support [DyHead](configs/dyhead) (#6823)\r\n- Support [ResNet Strikes Back](configs/resnet_strikes_back) (#7001)\r\n- Support [OpenImages Dataset](configs/openimages) (#6331)\r\n- Support [TIMM backbone](configs/timm_example) (#7020)\r\n- Support visualization for Panoptic Segmentation (#7041)\r\n\r\n### Bug Fixes\r\n\r\n- Fix bug for the best checkpoints can not be saved when the `key_score` is None (#7101)\r\n- Fix MixUp transform filter boxes failing case (#7080)\r\n- Add missing properties in SABLHead (#7091)\r\n- Fix bug when NaNs exist in confusion matrix (#7147)\r\n- Fix PALETTE AttributeError in downstream task (#7230)\r\n\r\n### Improvements\r\n\r\n- Speed up SimOTA matching (#7098)\r\n- Add Chinese translation of `docs_zh-CN/tutorials/init_cfg.md` (#7188)\r\n\r\n### Contributors\r\n\r\nA total of 20 developers contributed to this release.\r\nThanks @ZwwWayne, @hhaAndroid, @RangiLyu, @AronLin, @BIGWangYuDong, @jbwang1997, @zytx121, @chhluo, @shinya7y, @LuooChen, @dvansa, @siatwangmin, @del-zhenwu, @vikashranjan26, @haofanwang, @jamiechoi1995, @HJoonKwon, @yarkable, @zhijian-liu, @RangeKing\r\n\r\n## New Contributors\r\n* @LuooChen made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7101\r\n* @dvansa made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7080\r\n* @siatwangmin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6476\r\n* @vikashranjan26 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7091\r\n* @RangeKing made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7215\r\n* @zhijian-liu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7175\r\n* @HJoonKwon made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7187\r\n* @yarkable made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7188\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.21.0...v2.22.0","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Breaking Changes</h3><a id=\"user-content-breaking-changes\" class=\"anchor\" aria-label=\"Permalink: Breaking Changes\" href=\"#breaking-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>In order to support the visualization for Panoptic Segmentation, the <code>num_classes</code> can not be <code>None</code> when using the <code>get_palette</code> function to determine whether to use the panoptic palette.</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support MaskFormer: <a href=\"https://arxiv.org/abs/2107.06278\" rel=\"nofollow\">Per-Pixel Classification is Not All You Need for Semantic Segmentation</a> (#7212)</li>\n<li>Support DyHead: <a href=\"https://arxiv.org/abs/2106.08322\" rel=\"nofollow\">Dynamic Head: Unifying Object Detection Heads with Attentions</a> (#6823)</li>\n<li>Release a good recipe of using ResNet in object detectors pre-trained by <a href=\"https://arxiv.org/abs/2110.00476\" rel=\"nofollow\">ResNet Strikes Back</a>, which consistently brings about 3~4 mAP improvements over RetinaNet, Faster/Mask/Cascade Mask R-CNN (#7001)</li>\n<li>Support <a href=\"https://storage.googleapis.com/openimages/web/index.html\" rel=\"nofollow\">Open Images Dataset</a> (#6331)</li>\n<li>Support TIMM backbone: <a href=\"https://github.com/rwightman/pytorch-image-models\">PyTorch Image Models</a> (#7020)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"configs/maskformer\">MaskFormer</a> (#7212)</li>\n<li>Support <a href=\"configs/dyhead\">DyHead</a> (#6823)</li>\n<li>Support <a href=\"configs/resnet_strikes_back\">ResNet Strikes Back</a> (#7001)</li>\n<li>Support <a href=\"configs/openimages\">OpenImages Dataset</a> (#6331)</li>\n<li>Support <a href=\"configs/timm_example\">TIMM backbone</a> (#7020)</li>\n<li>Support visualization for Panoptic Segmentation (#7041)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix bug for the best checkpoints can not be saved when the <code>key_score</code> is None (#7101)</li>\n<li>Fix MixUp transform filter boxes failing case (#7080)</li>\n<li>Add missing properties in SABLHead (#7091)</li>\n<li>Fix bug when NaNs exist in confusion matrix (#7147)</li>\n<li>Fix PALETTE AttributeError in downstream task (#7230)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Speed up SimOTA matching (#7098)</li>\n<li>Add Chinese translation of <code>docs_zh-CN/tutorials/init_cfg.md</code> (#7188)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 20 developers contributed to this release.\nThanks @ZwwWayne, @hhaAndroid, @RangiLyu, @AronLin, @BIGWangYuDong, @jbwang1997, @zytx121, @chhluo, @shinya7y, @LuooChen, @dvansa, @siatwangmin, @del-zhenwu, @vikashranjan26, @haofanwang, @jamiechoi1995, @HJoonKwon, @yarkable, @zhijian-liu, @RangeKing</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@LuooChen made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7101\">https://github.com/open-mmlab/mmdetection/pull/7101</a>\n</li>\n<li>@dvansa made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7080\">https://github.com/open-mmlab/mmdetection/pull/7080</a>\n</li>\n<li>@siatwangmin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6476\">https://github.com/open-mmlab/mmdetection/pull/6476</a>\n</li>\n<li>@vikashranjan26 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7091\">https://github.com/open-mmlab/mmdetection/pull/7091</a>\n</li>\n<li>@RangeKing made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7215\">https://github.com/open-mmlab/mmdetection/pull/7215</a>\n</li>\n<li>@zhijian-liu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7175\">https://github.com/open-mmlab/mmdetection/pull/7175</a>\n</li>\n<li>@HJoonKwon made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7187\">https://github.com/open-mmlab/mmdetection/pull/7187</a>\n</li>\n<li>@yarkable made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7188\">https://github.com/open-mmlab/mmdetection/pull/7188</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.21.0...v2.22.0\">https://github.com/open-mmlab/mmdetection/compare/v2.21.0...v2.22.0</a></p>\n","assets":[]},{"id":58996004,"tag":"v2.21.0","author":"ZwwWayne","name":"MMDetection V2.21.0 Release","draft":false,"prerelease":false,"createdAt":"2022-02-08T10:54:03Z","publishedAt":"2022-02-08T10:56:07Z","markdown":"## Breaking Changes\r\n\r\nTo standardize the contents in config READMEs and meta files of OpenMMLab projects, the READMEs and meta files in each config directory have been significantly changed. The template will be released in the future, for now, you can refer to the examples of README for [algorithm](https://github.com/open-mmlab/mmdetection/blob/master/configs/faster_rcnn/README.md), [dataset](https://github.com/open-mmlab/mmdetection/blob/master/configs/deepfashion/README.md) and [backbone](https://github.com/open-mmlab/mmdetection/blob/master/configs/regnet/README.md). To align with the standard, the configs in dcn are put into to two directories named `dcn` and `dcnv2`.\r\n\r\n### New Features\r\n\r\n- Allow to customize colors of different classes during visualization (#6716)\r\n- Support CPU training (#7016)\r\n- Add download script of COCO, LVIS, and VOC dataset (#7015)\r\n\r\n#### Bug Fixes\r\n\r\n- Fix weight conversion issue of RetinaNet with Swin-S (#6973)\r\n- Update ``__repr__`` of ``Compose`` (#6951)\r\n- Fix BadZipFile Error when build docker (#6966)\r\n- Fix bug in non-distributed multi-gpu training/testing (#7019)\r\n- Fix bbox clamp in PyTorch 1.10 (#7074)\r\n- Relax the requirement of PALETTE in dataset wrappers (#7085)\r\n- Keep the same weights before reassign in the PAA head (#7032)\r\n- Update code demo in doc (#7092)\r\n\r\n### Improvements\r\n\r\n- Speed-up training by allow to set variables of multi-processing (#6974, #7036)\r\n- Add links of Chinese tutorials in readme (#6897)\r\n- Disable cv2 multiprocessing by default for acceleration (#6867)\r\n- Deprecate the support for \"python setup.py test\" (#6998)\r\n- Re-organize metafiles and config readmes (#7051)\r\n- Fix None grad problem during training TOOD by adding ``SigmoidGeometricMean`` (#7090)\r\n\r\n### Contributors\r\n\r\nA total of 26 developers contributed to this release.\r\nThanks @del-zhenwu, @zimoqingfeng, @srishilesh, @imyhxy, @jenhaoyang, @jliu-ac, @kimnamu, @ShengliLiu, @garvan2021, @ciusji, @DIYer22, @kimnamu, @q3394101, @zhouzaida, @gaotongxiao, @topsy404, @AntoAndGar, @jbwang1997, @nijkah, @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @AronLin\r\n\r\n## New Contributors\r\n* @srishilesh made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6936\r\n* @imyhxy made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6867\r\n* @jliu-ac made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6725\r\n* @ShengliLiu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6973\r\n* @garvan2021 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6951\r\n* @nijkah made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6956\r\n* @ciusji made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7033\r\n* @DIYer22 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6966\r\n* @topsy404 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7124\r\n* @AntoAndGar made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7095\r\n* @zimoqingfeng made their first contribution in https://github.com/open-mmlab/mmdetection/pull/7032\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.20.0...v2.21.0","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Breaking Changes</h2><a id=\"user-content-breaking-changes\" class=\"anchor\" aria-label=\"Permalink: Breaking Changes\" href=\"#breaking-changes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>To standardize the contents in config READMEs and meta files of OpenMMLab projects, the READMEs and meta files in each config directory have been significantly changed. The template will be released in the future, for now, you can refer to the examples of README for <a href=\"https://github.com/open-mmlab/mmdetection/blob/master/configs/faster_rcnn/README.md\">algorithm</a>, <a href=\"https://github.com/open-mmlab/mmdetection/blob/master/configs/deepfashion/README.md\">dataset</a> and <a href=\"https://github.com/open-mmlab/mmdetection/blob/master/configs/regnet/README.md\">backbone</a>. To align with the standard, the configs in dcn are put into to two directories named <code>dcn</code> and <code>dcnv2</code>.</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Allow to customize colors of different classes during visualization (#6716)</li>\n<li>Support CPU training (#7016)</li>\n<li>Add download script of COCO, LVIS, and VOC dataset (#7015)</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Bug Fixes</h4><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix weight conversion issue of RetinaNet with Swin-S (#6973)</li>\n<li>Update <code>__repr__</code> of <code>Compose</code> (#6951)</li>\n<li>Fix BadZipFile Error when build docker (#6966)</li>\n<li>Fix bug in non-distributed multi-gpu training/testing (#7019)</li>\n<li>Fix bbox clamp in PyTorch 1.10 (#7074)</li>\n<li>Relax the requirement of PALETTE in dataset wrappers (#7085)</li>\n<li>Keep the same weights before reassign in the PAA head (#7032)</li>\n<li>Update code demo in doc (#7092)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Speed-up training by allow to set variables of multi-processing (#6974, #7036)</li>\n<li>Add links of Chinese tutorials in readme (#6897)</li>\n<li>Disable cv2 multiprocessing by default for acceleration (#6867)</li>\n<li>Deprecate the support for \"python setup.py test\" (#6998)</li>\n<li>Re-organize metafiles and config readmes (#7051)</li>\n<li>Fix None grad problem during training TOOD by adding <code>SigmoidGeometricMean</code> (#7090)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 26 developers contributed to this release.\nThanks @del-zhenwu, @zimoqingfeng, @srishilesh, @imyhxy, @jenhaoyang, @jliu-ac, @kimnamu, @ShengliLiu, @garvan2021, @ciusji, @DIYer22, @kimnamu, @q3394101, @zhouzaida, @gaotongxiao, @topsy404, @AntoAndGar, @jbwang1997, @nijkah, @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @AronLin</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@srishilesh made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6936\">https://github.com/open-mmlab/mmdetection/pull/6936</a>\n</li>\n<li>@imyhxy made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6867\">https://github.com/open-mmlab/mmdetection/pull/6867</a>\n</li>\n<li>@jliu-ac made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6725\">https://github.com/open-mmlab/mmdetection/pull/6725</a>\n</li>\n<li>@ShengliLiu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6973\">https://github.com/open-mmlab/mmdetection/pull/6973</a>\n</li>\n<li>@garvan2021 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6951\">https://github.com/open-mmlab/mmdetection/pull/6951</a>\n</li>\n<li>@nijkah made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6956\">https://github.com/open-mmlab/mmdetection/pull/6956</a>\n</li>\n<li>@ciusji made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7033\">https://github.com/open-mmlab/mmdetection/pull/7033</a>\n</li>\n<li>@DIYer22 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6966\">https://github.com/open-mmlab/mmdetection/pull/6966</a>\n</li>\n<li>@topsy404 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7124\">https://github.com/open-mmlab/mmdetection/pull/7124</a>\n</li>\n<li>@AntoAndGar made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7095\">https://github.com/open-mmlab/mmdetection/pull/7095</a>\n</li>\n<li>@zimoqingfeng made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/7032\">https://github.com/open-mmlab/mmdetection/pull/7032</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.20.0...v2.21.0\">https://github.com/open-mmlab/mmdetection/compare/v2.20.0...v2.21.0</a></p>\n","assets":[]},{"id":56179087,"tag":"v2.20.0","author":"ZwwWayne","name":"MMDetection V2.20.0 Release","draft":false,"prerelease":false,"createdAt":"2021-12-30T09:22:44Z","publishedAt":"2021-12-30T09:27:38Z","markdown":"#### New Features\r\n\r\n- Support [TOOD](configs/tood/README.md): Task-aligned One-stage Object Detection (ICCV 2021 Oral) (#6746)\r\n- Support resuming from the latest checkpoint automatically (#6727)\r\n\r\n#### Bug Fixes\r\n\r\n- Fix wrong bbox `loss_weight` of the PAA head (#6744)\r\n- Fix the padding value of `gt_semantic_seg` in batch collating (#6837)\r\n- Fix test error of lvis when using `classwise` (#6845)\r\n- Avoid BC-breaking of `get_local_path`  (#6719)\r\n- Fix bug in `sync_norm_hook` when the BN layer does not exist (#6852)\r\n- Use pycocotools directly no matter what platform it is (#6838)\r\n\r\n#### Improvements\r\n\r\n- Add unit test for SimOTA with no valid bbox (#6770)\r\n- Use precommit to check readme (#6802)\r\n- Support selecting GPU-ids in non-distributed testing time (#6781)\r\n\r\n#### Contributors\r\n\r\nA total of 16 developers contributed to this release.\r\nThanks @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @jamiechoi1995, @AronLin, @Keiku, @gkagkos, @fcakyon, @www516717402, @vansin, @zactodd, @kimnamu, @jenhaoyang\r\n\r\n## New Contributors\r\n* @jamiechoi1995 made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6795\r\n* @Keiku made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6865\r\n* @gkagkos made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6744\r\n* @vansin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6858\r\n* @zactodd made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6764\r\n* @kimnamu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6906\r\n* @jenhaoyang made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6881\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.19.1...v2.20.0","html":"<div class=\"markdown-heading\"><h4 class=\"heading-element\">New Features</h4><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"configs/tood/README.md\">TOOD</a>: Task-aligned One-stage Object Detection (ICCV 2021 Oral) (#6746)</li>\n<li>Support resuming from the latest checkpoint automatically (#6727)</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Bug Fixes</h4><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix wrong bbox <code>loss_weight</code> of the PAA head (#6744)</li>\n<li>Fix the padding value of <code>gt_semantic_seg</code> in batch collating (#6837)</li>\n<li>Fix test error of lvis when using <code>classwise</code> (#6845)</li>\n<li>Avoid BC-breaking of <code>get_local_path</code>  (#6719)</li>\n<li>Fix bug in <code>sync_norm_hook</code> when the BN layer does not exist (#6852)</li>\n<li>Use pycocotools directly no matter what platform it is (#6838)</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Improvements</h4><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Add unit test for SimOTA with no valid bbox (#6770)</li>\n<li>Use precommit to check readme (#6802)</li>\n<li>Support selecting GPU-ids in non-distributed testing time (#6781)</li>\n</ul>\n<div class=\"markdown-heading\"><h4 class=\"heading-element\">Contributors</h4><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 16 developers contributed to this release.\nThanks @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @jamiechoi1995, @AronLin, @Keiku, @gkagkos, @fcakyon, @www516717402, @vansin, @zactodd, @kimnamu, @jenhaoyang</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@jamiechoi1995 made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6795\">https://github.com/open-mmlab/mmdetection/pull/6795</a>\n</li>\n<li>@Keiku made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6865\">https://github.com/open-mmlab/mmdetection/pull/6865</a>\n</li>\n<li>@gkagkos made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6744\">https://github.com/open-mmlab/mmdetection/pull/6744</a>\n</li>\n<li>@vansin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6858\">https://github.com/open-mmlab/mmdetection/pull/6858</a>\n</li>\n<li>@zactodd made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6764\">https://github.com/open-mmlab/mmdetection/pull/6764</a>\n</li>\n<li>@kimnamu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6906\">https://github.com/open-mmlab/mmdetection/pull/6906</a>\n</li>\n<li>@jenhaoyang made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6881\">https://github.com/open-mmlab/mmdetection/pull/6881</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.19.1...v2.20.0\">https://github.com/open-mmlab/mmdetection/compare/v2.19.1...v2.20.0</a></p>\n","assets":[]},{"id":55341802,"tag":"v2.19.1","author":"ZwwWayne","name":"MMDetection V2.19.1 Release","draft":false,"prerelease":false,"createdAt":"2021-12-15T06:55:50Z","publishedAt":"2021-12-15T08:10:45Z","markdown":"","html":"","assets":[]},{"id":54326773,"tag":"v2.19.0","author":"ZwwWayne","name":"MMDetection V2.19.0 Release","draft":false,"prerelease":false,"createdAt":"2021-11-30T14:13:41Z","publishedAt":"2021-11-30T14:15:59Z","markdown":"### Highlights\r\n\r\n- Support [Label Assignment Distillation](https://arxiv.org/abs/2108.10520)\r\n- Support `persistent_workers` for Pytorch >= 1.7\r\n- Align accuracy to the updated official YOLOX\r\n\r\n### New Features\r\n\r\n- Support [Label Assignment Distillation](https://arxiv.org/abs/2108.10520) (#6342)\r\n- Support `persistent_workers` for Pytorch >= 1.7 (#6435)\r\n\r\n### Bug Fixes\r\n\r\n- Fix repeatedly output warning message (#6584)\r\n- Avoid infinite GPU waiting in dist training (#6501)\r\n- Fix SSD512 config error (#6574)\r\n- Fix MMDetection model to ONNX command (#6558)\r\n\r\n### Improvements\r\n\r\n- Refactor configs of FP16 models (#6592)\r\n- Align accuracy to the updated official YOLOX (#6443)\r\n- Speed up training and reduce memory cost when using PhotoMetricDistortion. (#6442)\r\n- Make OHEM work with seesaw loss (#6514)\r\n\r\n\r\n### Documents\r\n\r\n- Update README.md (#6567)\r\n\r\n### Contributors\r\n\r\nA total of 11 developers contributed to this release.\r\nThanks @FloydHsiu, @RangiLyu, @ZwwWayne, @AndreaPi, @st9007a, @hachreak, @BIGWangYuDong, @hhaAndroid, @AronLin, @chhluo, @vealocia, @HarborYuan, @st9007a, @jshilong\r\n\r\n\r\n### New Contributors\r\n* @FloydHsiu made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6546\r\n","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2108.10520\" rel=\"nofollow\">Label Assignment Distillation</a>\n</li>\n<li>Support <code>persistent_workers</code> for Pytorch &gt;= 1.7</li>\n<li>Align accuracy to the updated official YOLOX</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"https://arxiv.org/abs/2108.10520\" rel=\"nofollow\">Label Assignment Distillation</a> (#6342)</li>\n<li>Support <code>persistent_workers</code> for Pytorch &gt;= 1.7 (#6435)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix repeatedly output warning message (#6584)</li>\n<li>Avoid infinite GPU waiting in dist training (#6501)</li>\n<li>Fix SSD512 config error (#6574)</li>\n<li>Fix MMDetection model to ONNX command (#6558)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Refactor configs of FP16 models (#6592)</li>\n<li>Align accuracy to the updated official YOLOX (#6443)</li>\n<li>Speed up training and reduce memory cost when using PhotoMetricDistortion. (#6442)</li>\n<li>Make OHEM work with seesaw loss (#6514)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Documents</h3><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Update README.md (#6567)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 11 developers contributed to this release.\nThanks @FloydHsiu, @RangiLyu, @ZwwWayne, @AndreaPi, @st9007a, @hachreak, @BIGWangYuDong, @hhaAndroid, @AronLin, @chhluo, @vealocia, @HarborYuan, @st9007a, @jshilong</p>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Contributors</h3><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@FloydHsiu made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6546\">https://github.com/open-mmlab/mmdetection/pull/6546</a>\n</li>\n</ul>\n","assets":[]},{"id":53410319,"tag":"v2.18.1","author":"ZwwWayne","name":"MMDetection V2.18.1 Release","draft":false,"prerelease":false,"createdAt":"2021-11-15T17:59:17Z","publishedAt":"2021-11-15T18:02:44Z","markdown":"### Highlights\r\n\r\n- Release [QueryInst](http://arxiv.org/abs/2105.01928) pre-trained weights (#6460)\r\n- Support plot confusion matrix (#6344)\r\n\r\n### New Features\r\n\r\n- Release [QueryInst](http://arxiv.org/abs/2105.01928) pre-trained weights (#6460)\r\n- Support plot confusion matrix (#6344)\r\n\r\n### Bug Fixes\r\n\r\n- Fix aug test error when the number of prediction bboxes is 0 (#6398)\r\n- Fix SpatialReductionAttention in PVT (#6488)\r\n- Fix wrong use of `trunc_normal_init` in PVT and Swin-Transformer (#6432)\r\n\r\n### Improvements\r\n\r\n- Save the printed AP information of COCO API to logger (#6505)\r\n- Always map location to cpu when load checkpoint (#6405)\r\n- Set a random seed when the user does not set a seed (#6457)\r\n\r\n### Documents\r\n\r\n- Chinese version of [Corruption Benchmarking](docs/robustness_benchmarking.mddocs/robustness_benchmarking.md) (#6375)\r\n- Fix config path in docs (#6396)\r\n- Update GRoIE readme (#6401)\r\n\r\n### Contributors\r\n\r\nA total of 11 developers contributed to this release.\r\nThanks @st9007a, @hachreak, @HarborYuan, @vealocia, @chhluo, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne\r\n\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.18.0...v2.18.1","html":"<div class=\"markdown-heading\"><h3 class=\"heading-element\">Highlights</h3><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Release <a href=\"http://arxiv.org/abs/2105.01928\" rel=\"nofollow\">QueryInst</a> pre-trained weights (#6460)</li>\n<li>Support plot confusion matrix (#6344)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">New Features</h3><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Release <a href=\"http://arxiv.org/abs/2105.01928\" rel=\"nofollow\">QueryInst</a> pre-trained weights (#6460)</li>\n<li>Support plot confusion matrix (#6344)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Bug Fixes</h3><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix aug test error when the number of prediction bboxes is 0 (#6398)</li>\n<li>Fix SpatialReductionAttention in PVT (#6488)</li>\n<li>Fix wrong use of <code>trunc_normal_init</code> in PVT and Swin-Transformer (#6432)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Improvements</h3><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Save the printed AP information of COCO API to logger (#6505)</li>\n<li>Always map location to cpu when load checkpoint (#6405)</li>\n<li>Set a random seed when the user does not set a seed (#6457)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Documents</h3><a id=\"user-content-documents\" class=\"anchor\" aria-label=\"Permalink: Documents\" href=\"#documents\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Chinese version of <a href=\"docs/robustness_benchmarking.mddocs/robustness_benchmarking.md\">Corruption Benchmarking</a> (#6375)</li>\n<li>Fix config path in docs (#6396)</li>\n<li>Update GRoIE readme (#6401)</li>\n</ul>\n<div class=\"markdown-heading\"><h3 class=\"heading-element\">Contributors</h3><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 11 developers contributed to this release.\nThanks @st9007a, @hachreak, @HarborYuan, @vealocia, @chhluo, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne</p>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.18.0...v2.18.1\">https://github.com/open-mmlab/mmdetection/compare/v2.18.0...v2.18.1</a></p>\n","assets":[]},{"id":52195111,"tag":"v2.18.0","author":"ZwwWayne","name":"MMDetection V2.18.0 Release","draft":false,"prerelease":false,"createdAt":"2021-10-28T02:17:17Z","publishedAt":"2021-10-28T02:31:37Z","markdown":"## Highlights\r\n\r\n- Support [QueryInst](http://arxiv.org/abs/2105.01928) (#6050)\r\n- Refactor dense heads to decouple onnx export logics from `get_bboxes` and speed up inference (#5317, #6003, #6369, #6268, #6315)\r\n\r\n## New Features\r\n\r\n- Support [QueryInst](http://arxiv.org/abs/2105.01928) (#6050)\r\n- Support infinite sampler (#5996)\r\n\r\n## Bug Fixes\r\n\r\n- Fix `init_weight` in fcn_mask_head (#6378)\r\n- Fix type error in `imshow_bboxes` of RPN (#6386)\r\n- Fix broken colab link in MMDetection Tutorial (#6382)\r\n- Make sure the device and dtype of `scale_factor` are the same as bboxes (#6374)\r\n- Remove sampling hardcode (#6317)\r\n- Fix `RandomAffine` bbox coordinate bug (#6293)\r\n- Fix initialization bug of final cls/reg layer in convfc head (#6279)\r\n- Fix `img_shape` broken in auto_augment (#6259)\r\n- Fix kwargs parameter missing error in two stage detector (#6256)\r\n\r\n## Improvements\r\n\r\n- Unify the interface of stuff head and panoptic head (#6308)\r\n- Polish readme (#6243)\r\n- Add code-spell pre-commit hook and fix a typo (#6306)\r\n- Fix typos (#6245, #6190)\r\n- Fix sampler unit test (#6284)\r\n- Fix `forward_dummy` of YOLACT to enable `get_flops` (#6079)\r\n- Fix link error in the config documentation (#6252)\r\n- Adjust the order to beautify the document (#6195)\r\n\r\n## Refactors\r\n\r\n- Refactor one-stage get_bboxes logic (#5317)\r\n- Refactor ONNX export of One-Stage models (#6003, #6369)\r\n- Refactor dense heads and speed-up (#6268)\r\n- Migrate to use prior_generator in the training of dense heads (#6315)\r\n\r\n## Contributors\r\n\r\nA total of 18 developers contributed to this release.\r\nThanks @Boyden, @onnkeat, @st9007a, @vealocia, @yhcao6, @DapangpangX, @yellowdolphin, @cclauss, @kennymckormick,\r\n@pingguokiller, @collinzrj, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @jshilong, @RangiLyu, @ZwwWayne\r\n\r\n## New Contributors\r\n* @AndreaPi made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6252\r\n* @pingguokiller made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6256\r\n* @kennymckormick made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6245\r\n* @yellowdolphin made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6259\r\n* @DapangpangX made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6293\r\n* @vealocia made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6050\r\n* @st9007a made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6374\r\n* @onnkeat made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6382\r\n* @Boyden made their first contribution in https://github.com/open-mmlab/mmdetection/pull/6378\r\n\r\n**Full Changelog**: https://github.com/open-mmlab/mmdetection/compare/v2.17.0...v2.18.0","html":"<div class=\"markdown-heading\"><h2 class=\"heading-element\">Highlights</h2><a id=\"user-content-highlights\" class=\"anchor\" aria-label=\"Permalink: Highlights\" href=\"#highlights\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"http://arxiv.org/abs/2105.01928\" rel=\"nofollow\">QueryInst</a> (#6050)</li>\n<li>Refactor dense heads to decouple onnx export logics from <code>get_bboxes</code> and speed up inference (#5317, #6003, #6369, #6268, #6315)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Features</h2><a id=\"user-content-new-features\" class=\"anchor\" aria-label=\"Permalink: New Features\" href=\"#new-features\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Support <a href=\"http://arxiv.org/abs/2105.01928\" rel=\"nofollow\">QueryInst</a> (#6050)</li>\n<li>Support infinite sampler (#5996)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Bug Fixes</h2><a id=\"user-content-bug-fixes\" class=\"anchor\" aria-label=\"Permalink: Bug Fixes\" href=\"#bug-fixes\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Fix <code>init_weight</code> in fcn_mask_head (#6378)</li>\n<li>Fix type error in <code>imshow_bboxes</code> of RPN (#6386)</li>\n<li>Fix broken colab link in MMDetection Tutorial (#6382)</li>\n<li>Make sure the device and dtype of <code>scale_factor</code> are the same as bboxes (#6374)</li>\n<li>Remove sampling hardcode (#6317)</li>\n<li>Fix <code>RandomAffine</code> bbox coordinate bug (#6293)</li>\n<li>Fix initialization bug of final cls/reg layer in convfc head (#6279)</li>\n<li>Fix <code>img_shape</code> broken in auto_augment (#6259)</li>\n<li>Fix kwargs parameter missing error in two stage detector (#6256)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Improvements</h2><a id=\"user-content-improvements\" class=\"anchor\" aria-label=\"Permalink: Improvements\" href=\"#improvements\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Unify the interface of stuff head and panoptic head (#6308)</li>\n<li>Polish readme (#6243)</li>\n<li>Add code-spell pre-commit hook and fix a typo (#6306)</li>\n<li>Fix typos (#6245, #6190)</li>\n<li>Fix sampler unit test (#6284)</li>\n<li>Fix <code>forward_dummy</code> of YOLACT to enable <code>get_flops</code> (#6079)</li>\n<li>Fix link error in the config documentation (#6252)</li>\n<li>Adjust the order to beautify the document (#6195)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Refactors</h2><a id=\"user-content-refactors\" class=\"anchor\" aria-label=\"Permalink: Refactors\" href=\"#refactors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>Refactor one-stage get_bboxes logic (#5317)</li>\n<li>Refactor ONNX export of One-Stage models (#6003, #6369)</li>\n<li>Refactor dense heads and speed-up (#6268)</li>\n<li>Migrate to use prior_generator in the training of dense heads (#6315)</li>\n</ul>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">Contributors</h2><a id=\"user-content-contributors\" class=\"anchor\" aria-label=\"Permalink: Contributors\" href=\"#contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<p>A total of 18 developers contributed to this release.\nThanks @Boyden, @onnkeat, @st9007a, @vealocia, @yhcao6, @DapangpangX, @yellowdolphin, @cclauss, @kennymckormick,\n@pingguokiller, @collinzrj, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @jshilong, @RangiLyu, @ZwwWayne</p>\n<div class=\"markdown-heading\"><h2 class=\"heading-element\">New Contributors</h2><a id=\"user-content-new-contributors\" class=\"anchor\" aria-label=\"Permalink: New Contributors\" href=\"#new-contributors\"><span aria-hidden=\"true\" class=\"octicon octicon-link\"></span></a></div>\n<ul>\n<li>@AndreaPi made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6252\">https://github.com/open-mmlab/mmdetection/pull/6252</a>\n</li>\n<li>@pingguokiller made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6256\">https://github.com/open-mmlab/mmdetection/pull/6256</a>\n</li>\n<li>@kennymckormick made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6245\">https://github.com/open-mmlab/mmdetection/pull/6245</a>\n</li>\n<li>@yellowdolphin made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6259\">https://github.com/open-mmlab/mmdetection/pull/6259</a>\n</li>\n<li>@DapangpangX made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6293\">https://github.com/open-mmlab/mmdetection/pull/6293</a>\n</li>\n<li>@vealocia made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6050\">https://github.com/open-mmlab/mmdetection/pull/6050</a>\n</li>\n<li>@st9007a made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6374\">https://github.com/open-mmlab/mmdetection/pull/6374</a>\n</li>\n<li>@onnkeat made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6382\">https://github.com/open-mmlab/mmdetection/pull/6382</a>\n</li>\n<li>@Boyden made their first contribution in <a href=\"https://github.com/open-mmlab/mmdetection/pull/6378\">https://github.com/open-mmlab/mmdetection/pull/6378</a>\n</li>\n</ul>\n<p><strong>Full Changelog</strong>: <a href=\"https://github.com/open-mmlab/mmdetection/compare/v2.17.0...v2.18.0\">https://github.com/open-mmlab/mmdetection/compare/v2.17.0...v2.18.0</a></p>\n","assets":[]}]}