Hi guys, \nI just want to let you know there is a new paper called Poly-YOLO. It shows that YOLOv3 includes two principal issues: a label can rewrite another one, and the anchors' distribution among output scales is not ideal. Because YOLOv4 uses the same neck and head as YOLOv3, it suffers from the two problems too. Poly-YOLO solves the issues by a new neck and head. It also proposes an additional (not necessary) improvement: a way to realize instance segmentation, but this is not the aim of this competition so it is useless here.
\n\nBecause Poly-YOLO increases the original YOLOv3 performance significantly notwithstanding it uses fewer parameters in the backbone, it would be beneficial for the competition to take the new neck+head and the things around and integrate them into YOLOv4. Then, the accuracy should be amazing (at least I hope). The links are below. As I am highly interested in Poly-YOLO, and I understand the principle (but I am not an expert in TensorFlow), do not hesitate to write to me for a more in-depth explanation if you do not understand the theory behind.
\n\nPaper on arXiv: https://arxiv.org/abs/2005.13243\nBlog on Medium: https://tinyurl.com/yd3bsrom\nGitlab: https://gitlab.com/irafm-ai/poly-yolo
", "messages": [ { "id": "867942", "postDate": "05/30/2020 18:52:14", "content": "Hi guys, \nI just want to let you know there is a new paper called Poly-YOLO. It shows that YOLOv3 includes two principal issues: a label can rewrite another one, and the anchors' distribution among output scales is not ideal. Because YOLOv4 uses the same neck and head as YOLOv3, it suffers from the two problems too. Poly-YOLO solves the issues by a new neck and head. It also proposes an additional (not necessary) improvement: a way to realize instance segmentation, but this is not the aim of this competition so it is useless here.
\n\nBecause Poly-YOLO increases the original YOLOv3 performance significantly notwithstanding it uses fewer parameters in the backbone, it would be beneficial for the competition to take the new neck+head and the things around and integrate them into YOLOv4. Then, the accuracy should be amazing (at least I hope). The links are below. As I am highly interested in Poly-YOLO, and I understand the principle (but I am not an expert in TensorFlow), do not hesitate to write to me for a more in-depth explanation if you do not understand the theory behind.
\n\nPaper on arXiv: https://arxiv.org/abs/2005.13243\nBlog on Medium: https://tinyurl.com/yd3bsrom\nGitlab: https://gitlab.com/irafm-ai/poly-yolo
", "rawMarkdown": "Hi guys, \nI just want to let you know there is a new paper called Poly-YOLO. It shows that YOLOv3 includes two principal issues: a label can rewrite another one, and the anchors' distribution among output scales is not ideal. Because YOLOv4 uses the same neck and head as YOLOv3, it suffers from the two problems too. Poly-YOLO solves the issues by a new neck and head. It also proposes an additional (not necessary) improvement: a way to realize instance segmentation, but this is not the aim of this competition so it is useless here.\n\nBecause Poly-YOLO increases the original YOLOv3 performance significantly notwithstanding it uses fewer parameters in the backbone, it would be beneficial for the competition to take the new neck+head and the things around and integrate them into YOLOv4. Then, the accuracy should be amazing (at least I hope). The links are below. As I am highly interested in Poly-YOLO, and I understand the principle (but I am not an expert in TensorFlow), do not hesitate to write to me for a more in-depth explanation if you do not understand the theory behind.\n\nPaper on arXiv: [https://arxiv.org/abs/2005.13243](https://arxiv.org/abs/2005.13243)\nBlog on Medium: [https://tinyurl.com/yd3bsrom](https://tinyurl.com/yd3bsrom)\nGitlab: [https://gitlab.com/irafm-ai/poly-yolo](https://gitlab.com/irafm-ai/poly-yolo)", "votes": null }, { "id": "868142", "postDate": "05/31/2020 01:07:24", "content": "of course very interesting!!! \nDid you try your model with the wheat dataset and does it work on tpu?
\n\nkeep rocking...
", "rawMarkdown": "of course very interesting!!! \nDid you try your model with the wheat dataset and does it work on tpu?\n\nkeep rocking...", "votes": null }, { "id": "868238", "postDate": "05/31/2020 04:53:08", "content": "No, unfortunately, I do not have space and free hardware to participate in the competition:( And to be honest, the TPU processing is far from my knowledge level. \nI just read the discussion here and saw that YOLOv4 was tested and its results are not great. So I took a look at the dataset and kernels that analyze the data and it seems that the wheat dataset is not ideal for YOLOv4 precisely because the label rewriting and anchors distribution problem.
", "rawMarkdown": "No, unfortunately, I do not have space and free hardware to participate in the competition:( And to be honest, the TPU processing is far from my knowledge level. \nI just read the discussion here and saw that YOLOv4 was tested and its results are not great. So I took a look at the dataset and kernels that analyze the data and it seems that the wheat dataset is not ideal for YOLOv4 precisely because the label rewriting and anchors distribution problem.", "votes": null }, { "id": "892855", "postDate": "06/19/2020 07:57:07", "content": "Hi @mastemace ,
\n\nI've read your recent paper and I think polo-yolo would be ideal for an image segmentation project I'm working on. Where can I find resources that'll guide me in implementing poly-yolo in a project?
", "rawMarkdown": "Hi @mastemace , \n\nI've read your recent paper and I think polo-yolo would be ideal for an image segmentation project I'm working on. Where can I find resources that'll guide me in implementing poly-yolo in a project?", "votes": null }, { "id": "892886", "postDate": "06/19/2020 08:26:50", "content": "Hi, \nsome details can be found in Gitlab \nThere are active maintainers so do not hesitate to open an issue and ask to everything
", "rawMarkdown": "Hi, \nsome details can be found in [Gitlab](https://gitlab.com/irafm-ai/poly-yolo) \nThere are active maintainers so do not hesitate to open an issue and ask to everything", "votes": null } ], "comments": [ { "id": 868142, "author_name": "romanweilguny", "author_url": "", "post_date": "05/31/2020 01:07:24", "content": "of course very interesting!!! \nDid you try your model with the wheat dataset and does it work on tpu?
\n\nkeep rocking...
", "votes": null, "replies": [ { "id": 868238, "author_name": "mastemace", "author_url": "", "post_date": "05/31/2020 04:53:08", "content": "No, unfortunately, I do not have space and free hardware to participate in the competition:( And to be honest, the TPU processing is far from my knowledge level. \nI just read the discussion here and saw that YOLOv4 was tested and its results are not great. So I took a look at the dataset and kernels that analyze the data and it seems that the wheat dataset is not ideal for YOLOv4 precisely because the label rewriting and anchors distribution problem.
", "votes": null, "replies": [] } ] }, { "id": 892855, "author_name": "chimosky", "author_url": "", "post_date": "06/19/2020 07:57:07", "content": "Hi @mastemace ,
\n\nI've read your recent paper and I think polo-yolo would be ideal for an image segmentation project I'm working on. Where can I find resources that'll guide me in implementing poly-yolo in a project?
", "votes": null, "replies": [ { "id": 892886, "author_name": "mastemace", "author_url": "", "post_date": "06/19/2020 08:26:50", "content": "Hi, \nsome details can be found in Gitlab \nThere are active maintainers so do not hesitate to open an issue and ask to everything
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "867942": "Hi guys, \nI just want to let you know there is a new paper called Poly-YOLO. It shows that YOLOv3 includes two principal issues: a label can rewrite another one, and the anchors' distribution among output scales is not ideal. Because YOLOv4 uses the same neck and head as YOLOv3, it suffers from the two problems too. Poly-YOLO solves the issues by a new neck and head. It also proposes an additional (not necessary) improvement: a way to realize instance segmentation, but this is not the aim of this competition so it is useless here.\n\nBecause Poly-YOLO increases the original YOLOv3 performance significantly notwithstanding it uses fewer parameters in the backbone, it would be beneficial for the competition to take the new neck+head and the things around and integrate them into YOLOv4. Then, the accuracy should be amazing (at least I hope). The links are below. As I am highly interested in Poly-YOLO, and I understand the principle (but I am not an expert in TensorFlow), do not hesitate to write to me for a more in-depth explanation if you do not understand the theory behind.\n\nPaper on arXiv: [https://arxiv.org/abs/2005.13243](https://arxiv.org/abs/2005.13243)\nBlog on Medium: [https://tinyurl.com/yd3bsrom](https://tinyurl.com/yd3bsrom)\nGitlab: [https://gitlab.com/irafm-ai/poly-yolo](https://gitlab.com/irafm-ai/poly-yolo)", "868142": "of course very interesting!!! \nDid you try your model with the wheat dataset and does it work on tpu?\n\nkeep rocking...", "868238": "No, unfortunately, I do not have space and free hardware to participate in the competition:( And to be honest, the TPU processing is far from my knowledge level. \nI just read the discussion here and saw that YOLOv4 was tested and its results are not great. So I took a look at the dataset and kernels that analyze the data and it seems that the wheat dataset is not ideal for YOLOv4 precisely because the label rewriting and anchors distribution problem.", "892855": "Hi @mastemace , \n\nI've read your recent paper and I think polo-yolo would be ideal for an image segmentation project I'm working on. Where can I find resources that'll guide me in implementing poly-yolo in a project?", "892886": "Hi, \nsome details can be found in [Gitlab](https://gitlab.com/irafm-ai/poly-yolo) \nThere are active maintainers so do not hesitate to open an issue and ask to everything" }, "source": "meta" }