![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F436969%2F44ea6fe6e61f22471be5788407518d7e%2FYOLOv4-performance.png?generation=1595469398768719&alt=media\")
Hi everyone. I was looking for the paper on YOLO-V5 but came across this, very useful link explaining how YOLO-V5 is worse than YOLO-V4 ? YOLO-V4 has more stable implementation and beats YOLO-V5. This is the post by the contributor himself.
\n\nhttps://github.com/pjreddie/darknet/issues/2198
", "messages": [ { "id": 921915, "postDate": "2020-07-09T16:45:23.540Z", "content": "Hi everyone. I was looking for the paper on YOLO-V5 but came across this, very useful link explaining how YOLO-V5 is worse than YOLO-V4 ? YOLO-V4 has more stable implementation and beats YOLO-V5. This is the post by the contributor himself.
\n\nhttps://github.com/pjreddie/darknet/issues/2198
", "rawMarkdown": "Hi everyone. I was looking for the paper on YOLO-V5 but came across this, very useful link explaining how YOLO-V5 is worse than YOLO-V4 ? YOLO-V4 has more stable implementation and beats YOLO-V5. This is the post by the contributor himself.\n\nhttps://github.com/pjreddie/darknet/issues/2198", "votes": 3 }, { "id": 940511, "postDate": "2020-07-23T02:01:22.957Z", "content": "YOLOv4 exceeds YOLOv5's performance on the COCO benchmark.\n
There are circumstances where these results may not generalize to a given use case, however. Consider a dataset where objects are in long, narrow rectangular boxes. The anchors should be updated. YOLOv5 automatically learns new anchors for training while YOLOv4 does not. One may consider this not part of the model architecture itself (as anchors could manually be retrained for YOLOv4), but this is largely semantics.
\n\nGenerally, users find the ease-of-use and fast training of YOLOv5 and out-of-the-box performance on any domain to be its most attractive features.
\n\nUltimately, both can be tried on the same problem, e.g. --\n- How to Train YOLO-v4\n- How to Train YOLO-v5
\n\nI've also written more extensively about YOLOv4 vs YOLOv5.
", "rawMarkdown": "YOLOv4 exceeds YOLOv5's performance on the COCO benchmark.\n\n\nThere are circumstances where these results may not generalize to a given use case, however. Consider a dataset where objects are in long, narrow rectangular boxes. The anchors should be updated. YOLOv5 automatically learns new anchors for training while YOLOv4 does not. One may consider this not part of the model architecture itself (as anchors could manually be retrained for YOLOv4), but this is largely semantics.\n\nGenerally, users find the ease-of-use and fast training of YOLOv5 and out-of-the-box performance on any domain to be its most attractive features.\n\nUltimately, both can be tried on the same problem, e.g. --\n- [How to Train YOLO-v4](https://blog.roboflow.ai/training-yolov4-on-a-custom-dataset/)\n- [How to Train YOLO-v5](https://blog.roboflow.ai/how-to-train-yolov5-on-a-custom-dataset/)\n\n\nI've also written more extensively about [YOLOv4 vs YOLOv5](https://blog.roboflow.ai/yolov4-versus-yolov5/).", "votes": 1 }, { "id": 923502, "postDate": "2020-07-10T23:34:59.407Z", "content": "Regardless of the conceptual arguments, I found latest CoCo reported mAP of Yolov5 better than V4 (Glenn still actively updates the V5 performance and options)
\n\nOn Wheat, V5 works out of the box, while it's still challenging for V4
", "rawMarkdown": "Regardless of the conceptual arguments, I found latest CoCo reported mAP of Yolov5 better than V4 (Glenn still actively updates the V5 performance and options)\n\nOn Wheat, V5 works out of the box, while it's still challenging for V4", "votes": 1, "replies": [ { "id": 923913, "postDate": "2020-07-11T07:06:42.890Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 923422, "postDate": "2020-07-10T20:34:40.410Z", "content": "Theres a paper? Could you link it?
", "rawMarkdown": "Theres a paper? Could you link it?", "replies": [ { "id": 923425, "postDate": "2020-07-10T20:41:22.273Z", "content": "It seems the problem is actually that there is no paper and they say it's still under active development. From what I understand, models are much faster but with a lower accuracy than Yolo V4.
", "rawMarkdown": "It seems the problem is actually that there is no paper and they say it's still under active development. From what I understand, models are much faster but with a lower accuracy than Yolo V4.", "votes": 1 }, { "id": 923441, "postDate": "2020-07-10T21:14:10.767Z", "content": "Oh, my bad. I misread his line that said \"I was looking for the paper on YOLO-V5\". I thought he said he was already looking at it. Thanks
", "rawMarkdown": "Oh, my bad. I misread his line that said \"I was looking **for** the paper on YOLO-V5\". I thought he said he was already looking at it. Thanks", "votes": 1 } ] }, { "id": 923412, "postDate": "2020-07-10T20:09:52.240Z", "content": "Thank you for pointing this out! I just found the same post when looking for \"yolo v5\" on Google to know more about this latest version... I'm totally new to object detection thus I wouldn't be able to answer your question.
\n\nHowever I would be interested in seeing the results of Yolo V4 in the current best scoring kernel, in order to open the discussion. @nvnnghia do you think it's possible to replace Yolo V5 with Yolo V4 in your work?
\n\nEDIT - More sources about the controversy:\n- https://news.ycombinator.com/item?id=23478151\n- https://blog.roboflow.ai/yolov4-versus-yolov5/
", "rawMarkdown": "Thank you for pointing this out! I just found the same post when looking for \"yolo v5\" on Google to know more about this latest version... I'm totally new to object detection thus I wouldn't be able to answer your question.\n\nHowever I would be interested in seeing the results of Yolo V4 in the [current best scoring kernel](https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling), in order to open the discussion. @nvnnghia do you think it's possible to replace Yolo V5 with Yolo V4 in your work?\n\nEDIT - More sources about the controversy:\n- https://news.ycombinator.com/item?id=23478151\n- https://blog.roboflow.ai/yolov4-versus-yolov5/" } ], "comments": [ { "id": 940511, "author_name": "joseph", "author_url": "", "post_date": "2020-07-23T02:01:22.957000", "content": "YOLOv4 exceeds YOLOv5's performance on the COCO benchmark.\n
There are circumstances where these results may not generalize to a given use case, however. Consider a dataset where objects are in long, narrow rectangular boxes. The anchors should be updated. YOLOv5 automatically learns new anchors for training while YOLOv4 does not. One may consider this not part of the model architecture itself (as anchors could manually be retrained for YOLOv4), but this is largely semantics.
\n\nGenerally, users find the ease-of-use and fast training of YOLOv5 and out-of-the-box performance on any domain to be its most attractive features.
\n\nUltimately, both can be tried on the same problem, e.g. --\n- How to Train YOLO-v4\n- How to Train YOLO-v5
\n\nI've also written more extensively about YOLOv4 vs YOLOv5.
", "votes": 1, "replies": [] }, { "id": 923502, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-07-10T23:34:59.407000", "content": "Regardless of the conceptual arguments, I found latest CoCo reported mAP of Yolov5 better than V4 (Glenn still actively updates the V5 performance and options)
\n\nOn Wheat, V5 works out of the box, while it's still challenging for V4
", "votes": 1, "replies": [ { "id": 923913, "author_name": "", "author_url": "", "post_date": "2020-07-11T07:06:42.890000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 923422, "author_name": "Matt", "author_url": "", "post_date": "2020-07-10T20:34:40.410000", "content": "Theres a paper? Could you link it?
", "votes": 0, "replies": [ { "id": 923425, "author_name": "Michaël Karpe", "author_url": "", "post_date": "2020-07-10T20:41:22.273000", "content": "It seems the problem is actually that there is no paper and they say it's still under active development. From what I understand, models are much faster but with a lower accuracy than Yolo V4.
", "votes": 1, "replies": [] }, { "id": 923441, "author_name": "Matt", "author_url": "", "post_date": "2020-07-10T21:14:10.767000", "content": "Oh, my bad. I misread his line that said \"I was looking for the paper on YOLO-V5\". I thought he said he was already looking at it. Thanks
", "votes": 1, "replies": [] } ] }, { "id": 923412, "author_name": "Michaël Karpe", "author_url": "", "post_date": "2020-07-10T20:09:52.240000", "content": "Thank you for pointing this out! I just found the same post when looking for \"yolo v5\" on Google to know more about this latest version... I'm totally new to object detection thus I wouldn't be able to answer your question.
\n\nHowever I would be interested in seeing the results of Yolo V4 in the current best scoring kernel, in order to open the discussion. @nvnnghia do you think it's possible to replace Yolo V5 with Yolo V4 in your work?
\n\nEDIT - More sources about the controversy:\n- https://news.ycombinator.com/item?id=23478151\n- https://blog.roboflow.ai/yolov4-versus-yolov5/
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "921915": "Hi everyone. I was looking for the paper on YOLO-V5 but came across this, very useful link explaining how YOLO-V5 is worse than YOLO-V4 ? YOLO-V4 has more stable implementation and beats YOLO-V5. This is the post by the contributor himself.\n\nhttps://github.com/pjreddie/darknet/issues/2198", "940511": "YOLOv4 exceeds YOLOv5's performance on the COCO benchmark.\n\n\nThere are circumstances where these results may not generalize to a given use case, however. Consider a dataset where objects are in long, narrow rectangular boxes. The anchors should be updated. YOLOv5 automatically learns new anchors for training while YOLOv4 does not. One may consider this not part of the model architecture itself (as anchors could manually be retrained for YOLOv4), but this is largely semantics.\n\nGenerally, users find the ease-of-use and fast training of YOLOv5 and out-of-the-box performance on any domain to be its most attractive features.\n\nUltimately, both can be tried on the same problem, e.g. --\n- [How to Train YOLO-v4](https://blog.roboflow.ai/training-yolov4-on-a-custom-dataset/)\n- [How to Train YOLO-v5](https://blog.roboflow.ai/how-to-train-yolov5-on-a-custom-dataset/)\n\n\nI've also written more extensively about [YOLOv4 vs YOLOv5](https://blog.roboflow.ai/yolov4-versus-yolov5/).", "923502": "Regardless of the conceptual arguments, I found latest CoCo reported mAP of Yolov5 better than V4 (Glenn still actively updates the V5 performance and options)\n\nOn Wheat, V5 works out of the box, while it's still challenging for V4", "923422": "Theres a paper? Could you link it?", "923412": "Thank you for pointing this out! I just found the same post when looking for \"yolo v5\" on Google to know more about this latest version... I'm totally new to object detection thus I wouldn't be able to answer your question.\n\nHowever I would be interested in seeing the results of Yolo V4 in the [current best scoring kernel](https://www.kaggle.com/nvnnghia/yolov5-pseudo-labeling), in order to open the discussion. @nvnnghia do you think it's possible to replace Yolo V5 with Yolo V4 in your work?\n\nEDIT - More sources about the controversy:\n- https://news.ycombinator.com/item?id=23478151\n- https://blog.roboflow.ai/yolov4-versus-yolov5/" } }