Hello, a new paper called DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution as achieved new SOTA results on COCO.
\n\nPytorch implementation available here: github.com/joe-siyuan-qiao/DetectoRS with models for ResNet-50 and ResNeXt-101-32x4d backbone.
Did someone already try it on the wheat dataset ?\nI will try it this week I have some time
\n\nAbstract:
\n\n> Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-theart 54.7% box AP for object detection, 47.1% mask AP for instance segmentation, and 49.6% PQ for panoptic segmentation.
", "messages": [ { "id": "875899", "postDate": "06/06/2020 08:46:12", "content": "Hello, a new paper called DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution as achieved new SOTA results on COCO.
\n\nPytorch implementation available here: github.com/joe-siyuan-qiao/DetectoRS with models for ResNet-50 and ResNeXt-101-32x4d backbone.
Did someone already try it on the wheat dataset ?\nI will try it this week I have some time
\n\nAbstract:
\n\n> Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-theart 54.7% box AP for object detection, 47.1% mask AP for instance segmentation, and 49.6% PQ for panoptic segmentation.
", "rawMarkdown": "Hello, a new paper called [DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution](https://arxiv.org/pdf/2006.02334v1.pdf) as achieved new SOTA results on COCO. \n\nPytorch implementation available here: [github.com/joe-siyuan-qiao/DetectoRS](https://github.com/joe-siyuan-qiao/DetectoRS) with models for `ResNet-50` and `ResNeXt-101-32x4d` backbone.\n\nDid someone already try it on the wheat dataset ?\nI will try it this week I have some time\n\nAbstract:\n\n> Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-theart 54.7% box AP for object detection, 47.1% mask AP for instance segmentation, and 49.6% PQ for panoptic segmentation.", "votes": null }, { "id": "875900", "postDate": "06/06/2020 08:48:38", "content": "Interesting! If I try it, I will report back
", "rawMarkdown": "Interesting! If I try it, I will report back", "votes": null }, { "id": "875902", "postDate": "06/06/2020 08:50:40", "content": "That would be great @matthewmasters, thanks !
", "rawMarkdown": "That would be great @matthewmasters, thanks !", "votes": null }, { "id": "877281", "postDate": "06/07/2020 13:10:56", "content": "R50 get 0.7281.
", "rawMarkdown": "R50 get 0.7281.", "votes": null }, { "id": "877332", "postDate": "06/07/2020 14:09:56", "content": "So, it is better than Efficient Det, right?
", "rawMarkdown": "So, it is better than Efficient Det, right?", "votes": null }, { "id": "877367", "postDate": "06/07/2020 14:37:27", "content": "oh wow, thanks for sharing. I'm assuming this is with augmentations and single model, right?
", "rawMarkdown": "oh wow, thanks for sharing. I'm assuming this is with augmentations and single model, right?", "votes": null }, { "id": "878000", "postDate": "06/08/2020 07:21:06", "content": "Thanks a lot for this results! We definitely have to dive into it!
", "rawMarkdown": "Thanks a lot for this results! We definitely have to dive into it!", "votes": null }, { "id": "878097", "postDate": "06/08/2020 08:46:26", "content": "@matthewmasters Yeah,just single model.
", "rawMarkdown": "matthewmasters Yeah,just single model.", "votes": null }, { "id": "878574", "postDate": "06/08/2020 16:36:03", "content": "Great, thanks for sharing. I've been trying to get it setup but been having some issues since I'm not familiar with mmdet. Hopefully I can get it working soon as it seems promising
", "rawMarkdown": "Great, thanks for sharing. I've been trying to get it setup but been having some issues since I'm not familiar with mmdet. Hopefully I can get it working soon as it seems promising", "votes": null }, { "id": "884531", "postDate": "06/13/2020 12:38:22", "content": "Hi guys, how do you handle the config file removing the segmentation part?
", "rawMarkdown": "Hi guys, how do you handle the config file removing the segmentation part?", "votes": null }, { "id": "884645", "postDate": "06/13/2020 14:17:05", "content": "mmdetection has an example of removing branches via config file. Please refer to information related to HTC in the official repo for details.
", "rawMarkdown": "mmdetection has an example of removing branches via config file. Please refer to information related to HTC in the official repo for details.", "votes": null }, { "id": "884649", "postDate": "06/13/2020 14:18:36", "content": "The problem I've encountered during training with mmdetection is with the nms/soft-nms thresholding values, which gave me significantly worse results when compared to other pipelines.
", "rawMarkdown": "The problem I've encountered during training with mmdetection is with the nms/soft-nms thresholding values, which gave me significantly worse results when compared to other pipelines.", "votes": null }, { "id": "888113", "postDate": "06/16/2020 06:23:36", "content": "Is there any baseline available for this model? I'm a bit new to such tasks and struggling in getting it to work. Thanks!
", "rawMarkdown": "Is there any baseline available for this model? I'm a bit new to such tasks and struggling in getting it to work. Thanks!", "votes": null }, { "id": "888140", "postDate": "06/16/2020 06:48:30", "content": "Would also love to see a baseline of this. I've put it on the backburner for now
", "rawMarkdown": "Would also love to see a baseline of this. I've put it on the backburner for now", "votes": null }, { "id": "888906", "postDate": "06/16/2020 16:31:01", "content": "@spytensor single model without tta and other post-process trick?👍
", "rawMarkdown": "spytensor single model without tta and other post-process trick?👍", "votes": null } ], "comments": [ { "id": 875900, "author_name": "matthewmasters", "author_url": "", "post_date": "06/06/2020 08:48:38", "content": "Interesting! If I try it, I will report back
", "votes": null, "replies": [ { "id": 875902, "author_name": "yohann84l", "author_url": "", "post_date": "06/06/2020 08:50:40", "content": "That would be great @matthewmasters, thanks !
", "votes": null, "replies": [] } ] }, { "id": 877281, "author_name": "spytensor", "author_url": "", "post_date": "06/07/2020 13:10:56", "content": "R50 get 0.7281.
", "votes": null, "replies": [ { "id": 877332, "author_name": "sergeyzlobin", "author_url": "", "post_date": "06/07/2020 14:09:56", "content": "So, it is better than Efficient Det, right?
", "votes": null, "replies": [] }, { "id": 877367, "author_name": "matthewmasters", "author_url": "", "post_date": "06/07/2020 14:37:27", "content": "oh wow, thanks for sharing. I'm assuming this is with augmentations and single model, right?
", "votes": null, "replies": [] }, { "id": 878000, "author_name": "yohann84l", "author_url": "", "post_date": "06/08/2020 07:21:06", "content": "Thanks a lot for this results! We definitely have to dive into it!
", "votes": null, "replies": [] }, { "id": 878097, "author_name": "spytensor", "author_url": "", "post_date": "06/08/2020 08:46:26", "content": "@matthewmasters Yeah,just single model.
", "votes": null, "replies": [] }, { "id": 878574, "author_name": "matthewmasters", "author_url": "", "post_date": "06/08/2020 16:36:03", "content": "Great, thanks for sharing. I've been trying to get it setup but been having some issues since I'm not familiar with mmdet. Hopefully I can get it working soon as it seems promising
", "votes": null, "replies": [] }, { "id": 888906, "author_name": "cswwp347724", "author_url": "", "post_date": "06/16/2020 16:31:01", "content": "@spytensor single model without tta and other post-process trick?👍
", "votes": null, "replies": [] } ] }, { "id": 884531, "author_name": "jqeric", "author_url": "", "post_date": "06/13/2020 12:38:22", "content": "Hi guys, how do you handle the config file removing the segmentation part?
", "votes": null, "replies": [ { "id": 884645, "author_name": "alexanderliao", "author_url": "", "post_date": "06/13/2020 14:17:05", "content": "mmdetection has an example of removing branches via config file. Please refer to information related to HTC in the official repo for details.
", "votes": null, "replies": [] } ] }, { "id": 884649, "author_name": "alexanderliao", "author_url": "", "post_date": "06/13/2020 14:18:36", "content": "The problem I've encountered during training with mmdetection is with the nms/soft-nms thresholding values, which gave me significantly worse results when compared to other pipelines.
", "votes": null, "replies": [] }, { "id": 888113, "author_name": "kaushal2896", "author_url": "", "post_date": "06/16/2020 06:23:36", "content": "Is there any baseline available for this model? I'm a bit new to such tasks and struggling in getting it to work. Thanks!
", "votes": null, "replies": [ { "id": 888140, "author_name": "matthewmasters", "author_url": "", "post_date": "06/16/2020 06:48:30", "content": "Would also love to see a baseline of this. I've put it on the backburner for now
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "875899": "Hello, a new paper called [DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution](https://arxiv.org/pdf/2006.02334v1.pdf) as achieved new SOTA results on COCO. \n\nPytorch implementation available here: [github.com/joe-siyuan-qiao/DetectoRS](https://github.com/joe-siyuan-qiao/DetectoRS) with models for `ResNet-50` and `ResNeXt-101-32x4d` backbone.\n\nDid someone already try it on the wheat dataset ?\nI will try it this week I have some time\n\nAbstract:\n\n> Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-theart 54.7% box AP for object detection, 47.1% mask AP for instance segmentation, and 49.6% PQ for panoptic segmentation.", "875900": "Interesting! If I try it, I will report back", "875902": "That would be great @matthewmasters, thanks !", "877281": "R50 get 0.7281.", "877332": "So, it is better than Efficient Det, right?", "877367": "oh wow, thanks for sharing. I'm assuming this is with augmentations and single model, right?", "878000": "Thanks a lot for this results! We definitely have to dive into it!", "878097": "matthewmasters Yeah,just single model.", "878574": "Great, thanks for sharing. I've been trying to get it setup but been having some issues since I'm not familiar with mmdet. Hopefully I can get it working soon as it seems promising", "884531": "Hi guys, how do you handle the config file removing the segmentation part?", "884645": "mmdetection has an example of removing branches via config file. Please refer to information related to HTC in the official repo for details.", "884649": "The problem I've encountered during training with mmdetection is with the nms/soft-nms thresholding values, which gave me significantly worse results when compared to other pipelines.", "888113": "Is there any baseline available for this model? I'm a bit new to such tasks and struggling in getting it to work. Thanks!", "888140": "Would also love to see a baseline of this. I've put it on the backburner for now", "888906": "spytensor single model without tta and other post-process trick?👍" }, "source": "meta" }