![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3508221%2F9ad01bd27d933789643f7b4916b7db71%2F2020-05-29%20122720.png?generation=1590722864562525&alt=media\")
Facebook has published a paper with a new approach to object detection task - using Transformers.
\n\nhttps://arxiv.org/abs/2005.12872
\n\n\nWe present a new method that views object detection as a direct set prediction problem. Our approach streamlines the detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive baselines.\n
The code is available here: https://github.com/facebookresearch/detr
", "messages": [ { "id": "863091", "postDate": "05/27/2020 04:10:38", "content": "Facebook has published a paper with a new approach to object detection task - using Transformers.
\n\nhttps://arxiv.org/abs/2005.12872
\n\n\nWe present a new method that views object detection as a direct set prediction problem. Our approach streamlines the detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive baselines.\n
The code is available here: https://github.com/facebookresearch/detr
", "rawMarkdown": "Facebook has published a paper with a new approach to object detection task - using Transformers.\n\nhttps://arxiv.org/abs/2005.12872\n\n```\nWe present a new method that views object detection as a direct set prediction problem. Our approach streamlines the detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive baselines.\n```\n\nThe code is available here: https://github.com/facebookresearch/detr", "votes": null }, { "id": "863286", "postDate": "05/27/2020 07:53:04", "content": "thanks for sharing.\nBTW, my favorite sentence in the paper :)
\n\n\n", "rawMarkdown": "thanks for sharing.\nBTW, my favorite sentence in the paper :)\n> Training the **baseline** model for 300 epochs on **16 V100 GPUs takes 3 days**", "votes": null }, { "id": "863416", "postDate": "05/27/2020 10:03:49", "content": "Training the baseline model for 300 epochs on 16 V100 GPUs takes 3 days
\n
I suppose they can afford the approach of grow big or go home :)
Before it was about a novel approach and elegant code.
\n\nToday it's basically \"MY MODEL TRAINED FOR 5 DAYS ON 100 GPUs BEAT THAT GOOGLE.\"
", "rawMarkdown": "Before it was about a novel approach and elegant code.\n\nToday it's basically \"MY MODEL TRAINED FOR 5 DAYS ON 100 GPUs BEAT THAT GOOGLE.\"", "votes": null }, { "id": "864308", "postDate": "05/28/2020 00:12:57", "content": "😂
", "rawMarkdown": "😂", "votes": null }, { "id": "865011", "postDate": "05/28/2020 10:30:32", "content": "@artgor my bad, I didn't read the headline of your post. I found it today's morning meeting in my office and shared it here as hope someone may find it useful but some people are just downvoting until @matthewmasters informed me about the cause, though I didn't follow my post after sharing. 😂
\n\nHowever, Thanks @artgor for this information and also @matthewmasters for letting me know. 😃
", "rawMarkdown": "artgor my bad, I didn't read the headline of your post. I found it today's morning meeting in my office and shared it here as hope someone may find it useful but some people are just downvoting until @matthewmasters informed me about the cause, though I didn't follow my post after sharing. 😂 \n\nHowever, Thanks @artgor for this information and also @matthewmasters for letting me know. 😃", "votes": null }, { "id": "865271", "postDate": "05/28/2020 14:11:45", "content": "No problem :)
", "rawMarkdown": "No problem :)", "votes": null }, { "id": "866014", "postDate": "05/29/2020 03:27:49", "content": "Quite interesting, thanks for sharing!
\n\nDo you guys think it will perform well on the wheat dataset?\nI'm quite skeptical because DETR performs worse on small objects (APs) compared to FRCNNs. (It does great for large objects) However, it might do fine for detecting partially-occluded objects, since Transformers will find its relationships.
\n\n
I have try out DETR today with about 40 map@0.5:0.75. It's far worse than FasterRCNN. Maybe I still need tune hyperparameter and train more epochs.
", "rawMarkdown": "I have try out DETR today with about 40 map@0.5:0.75. It's far worse than FasterRCNN. Maybe I still need tune hyperparameter and train more epochs.", "votes": null }, { "id": "866607", "postDate": "05/29/2020 14:39:06", "content": "How long did you train for? It seems to me this method has extremely high computational requirements and would probably take a long time to train.
", "rawMarkdown": "How long did you train for? It seems to me this method has extremely high computational requirements and would probably take a long time to train.", "votes": null }, { "id": "873731", "postDate": "06/04/2020 11:42:46", "content": "can you explain more or making videos on youtube for making the issues easy
", "rawMarkdown": "can you explain more or making videos on youtube for making the issues easy", "votes": null } ], "comments": [ { "id": 863286, "author_name": "valanm", "author_url": "", "post_date": "05/27/2020 07:53:04", "content": "thanks for sharing.\nBTW, my favorite sentence in the paper :)
\n\n\n", "votes": null, "replies": [ { "id": 863416, "author_name": "artgor", "author_url": "", "post_date": "05/27/2020 10:03:49", "content": "Training the baseline model for 300 epochs on 16 V100 GPUs takes 3 days
\n
I suppose they can afford the approach of grow big or go home :)
Before it was about a novel approach and elegant code.
\n\nToday it's basically \"MY MODEL TRAINED FOR 5 DAYS ON 100 GPUs BEAT THAT GOOGLE.\"
", "votes": null, "replies": [] }, { "id": 864308, "author_name": "xujingzhao", "author_url": "", "post_date": "05/28/2020 00:12:57", "content": "😂
", "votes": null, "replies": [] } ] }, { "id": 865011, "author_name": "ipythonx", "author_url": "", "post_date": "05/28/2020 10:30:32", "content": "@artgor my bad, I didn't read the headline of your post. I found it today's morning meeting in my office and shared it here as hope someone may find it useful but some people are just downvoting until @matthewmasters informed me about the cause, though I didn't follow my post after sharing. 😂
\n\nHowever, Thanks @artgor for this information and also @matthewmasters for letting me know. 😃
", "votes": null, "replies": [ { "id": 865271, "author_name": "artgor", "author_url": "", "post_date": "05/28/2020 14:11:45", "content": "No problem :)
", "votes": null, "replies": [] } ] }, { "id": 866014, "author_name": "kyoshioka47", "author_url": "", "post_date": "05/29/2020 03:27:49", "content": "Quite interesting, thanks for sharing!
\n\nDo you guys think it will perform well on the wheat dataset?\nI'm quite skeptical because DETR performs worse on small objects (APs) compared to FRCNNs. (It does great for large objects) However, it might do fine for detecting partially-occluded objects, since Transformers will find its relationships.
\n\n
I have try out DETR today with about 40 map@0.5:0.75. It's far worse than FasterRCNN. Maybe I still need tune hyperparameter and train more epochs.
", "votes": null, "replies": [] }, { "id": 866607, "author_name": "matthewmasters", "author_url": "", "post_date": "05/29/2020 14:39:06", "content": "How long did you train for? It seems to me this method has extremely high computational requirements and would probably take a long time to train.
", "votes": null, "replies": [] } ] }, { "id": 873731, "author_name": "esswidi125", "author_url": "", "post_date": "06/04/2020 11:42:46", "content": "can you explain more or making videos on youtube for making the issues easy
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "863091": "Facebook has published a paper with a new approach to object detection task - using Transformers.\n\nhttps://arxiv.org/abs/2005.12872\n\n```\nWe present a new method that views object detection as a direct set prediction problem. Our approach streamlines the detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive baselines.\n```\n\nThe code is available here: https://github.com/facebookresearch/detr", "863286": "thanks for sharing.\nBTW, my favorite sentence in the paper :)\n> Training the **baseline** model for 300 epochs on **16 V100 GPUs takes 3 days**", "863416": "I suppose they can afford the approach of `grow big or go home` :)", "863625": "Before it was about a novel approach and elegant code.\n\nToday it's basically \"MY MODEL TRAINED FOR 5 DAYS ON 100 GPUs BEAT THAT GOOGLE.\"", "864308": "😂", "865011": "artgor my bad, I didn't read the headline of your post. I found it today's morning meeting in my office and shared it here as hope someone may find it useful but some people are just downvoting until @matthewmasters informed me about the cause, though I didn't follow my post after sharing. 😂 \n\nHowever, Thanks @artgor for this information and also @matthewmasters for letting me know. 😃", "865271": "No problem :)", "866014": "Quite interesting, thanks for sharing! \n\nDo you guys think it will perform well on the wheat dataset?\nI'm quite skeptical because DETR performs worse on small objects (APs) compared to FRCNNs. (It does great for large objects) However, it might do fine for detecting partially-occluded objects, since Transformers will find its relationships.\n\n", "866546": "I have try out DETR today with about 40 map@0.5:0.75. It's far worse than FasterRCNN. Maybe I still need tune hyperparameter and train more epochs.", "866607": "How long did you train for? It seems to me this method has extremely high computational requirements and would probably take a long time to train.", "873731": "can you explain more or making videos on youtube for making the issues easy" }, "source": "meta" }