{ "id": 175679, "title": "Yolov5 submissions removed?", "url": "/competitions/global-wheat-detection/discussion/175679", "author_name": "", "post_date": "2020-08-19T03:37:30.174626100Z", "votes": 6, "comment_count": 4, "views": 0, "content": "

There's a huge change in the leaderboard and it seems that submissions have been checked.
\nIt must have taken a lot of effort.Thanks so much to the Kaggle and University of Saskatchewan teams.

\n

No matter what the final private leaderboard is, we will be very satisfied because we have learned a lot from this competition:).Expecially thanks to the people who share any idea in the forum!

\n

Our team has trained three model:faster-rcnn,EfficientDet,DETR. And we used the fine-tune model to fix the training data.The faster-rcnn+EfficientDet merge score is 0.7649.
\nWe will also share our detail solution soon.
\nBtw, will the result be rerun again?

", "messages": [ { "id": "976700", "postDate": "08/19/2020 03:37:30", "content": "

There's a huge change in the leaderboard and it seems that submissions have been checked.
\nIt must have taken a lot of effort.Thanks so much to the Kaggle and University of Saskatchewan teams.

\n

No matter what the final private leaderboard is, we will be very satisfied because we have learned a lot from this competition:).Expecially thanks to the people who share any idea in the forum!

\n

Our team has trained three model:faster-rcnn,EfficientDet,DETR. And we used the fine-tune model to fix the training data.The faster-rcnn+EfficientDet merge score is 0.7649.
\nWe will also share our detail solution soon.
\nBtw, will the result be rerun again?

", "rawMarkdown": "There's a huge change in the leaderboard and it seems that submissions have been checked.\nIt must have taken a lot of effort.Thanks so much to the Kaggle and University of Saskatchewan teams.\n\nNo matter what the final private leaderboard is, we will be very satisfied because we have learned a lot from this competition:).Expecially thanks to the people who share any idea in the forum!\n\nOur team has trained three model:faster-rcnn,EfficientDet,DETR. And we used the fine-tune model to fix the training data.The faster-rcnn+EfficientDet merge score is 0.7649.\nWe will also share our detail solution soon.\n**Btw, will the result be rerun again?**", "votes": null }, { "id": "976984", "postDate": "08/19/2020 08:03:34", "content": "

Thanks for sharing @li325040229

\n

Could you explain DETR more please?
\nI also tried to train DETR model but lb is not good.

", "rawMarkdown": "Thanks for sharing @li325040229 \n\nCould you explain DETR more please?\nI also tried to train DETR model but lb is not good.", "votes": null }, { "id": "977056", "postDate": "08/19/2020 09:20:10", "content": "

Hi, I am his teammate who worked on DETR. I remember the single model of DETR(no PL, simple TTA WBF) almost (but not) passed 0.7 on the former public set. What about yours?

", "rawMarkdown": "Hi, I am his teammate who worked on DETR. I remember the single model of DETR(no PL, simple TTA WBF) almost (but not) passed 0.7 on the former public set. What about yours?", "votes": null }, { "id": "977061", "postDate": "08/19/2020 09:25:53", "content": "

Thanks for reply.
\nI commented this.
\nWhen i got this scores, i decided to choose good baseline model(e.g. EfficientDet, FasterRCNN, etc.).

", "rawMarkdown": "Thanks for reply.\nI commented [this](https://www.kaggle.com/c/global-wheat-detection/discussion/167092).\nWhen i got this scores, i decided to choose good baseline model(e.g. EfficientDet, FasterRCNN, etc.).", "votes": null }, { "id": "977134", "postDate": "08/19/2020 10:08:37", "content": "

My training script was mainly forked from [https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr]
\nIndeed DETR doesn't beat EfficientDet or FasterRCNN which can easily achieve 0.74+. It is still a feasible method. I will test it more in other contests. I tried adding FPN or using other backbones, but I am not 100% sure these methods make it outperform the original one. As now the public scores except the final two are hidden. I can't remember which version of DETR achieved the best score. If you are interested, you can view my inference script at all versions here[https://www.kaggle.com/qingyuanwang/detr-inference]

", "rawMarkdown": "My training script was mainly forked from [https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr]\nIndeed DETR doesn't beat EfficientDet or FasterRCNN which can easily achieve 0.74+. It is still a feasible method. I will test it more in other contests. I tried adding FPN or using other backbones, but I am not 100% sure these methods make it outperform the original one. As now the public scores except the final two are hidden. I can't remember which version of DETR achieved the best score. If you are interested, you can view my inference script at all versions here[https://www.kaggle.com/qingyuanwang/detr-inference]", "votes": null } ], "comments": [ { "id": 976984, "author_name": "piantic", "author_url": "", "post_date": "08/19/2020 08:03:34", "content": "

Thanks for sharing @li325040229

\n

Could you explain DETR more please?
\nI also tried to train DETR model but lb is not good.

", "votes": null, "replies": [ { "id": 977056, "author_name": "qingyuanwang", "author_url": "", "post_date": "08/19/2020 09:20:10", "content": "

Hi, I am his teammate who worked on DETR. I remember the single model of DETR(no PL, simple TTA WBF) almost (but not) passed 0.7 on the former public set. What about yours?

", "votes": null, "replies": [] }, { "id": 977061, "author_name": "piantic", "author_url": "", "post_date": "08/19/2020 09:25:53", "content": "

Thanks for reply.
\nI commented this.
\nWhen i got this scores, i decided to choose good baseline model(e.g. EfficientDet, FasterRCNN, etc.).

", "votes": null, "replies": [] }, { "id": 977134, "author_name": "qingyuanwang", "author_url": "", "post_date": "08/19/2020 10:08:37", "content": "

My training script was mainly forked from [https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr]
\nIndeed DETR doesn't beat EfficientDet or FasterRCNN which can easily achieve 0.74+. It is still a feasible method. I will test it more in other contests. I tried adding FPN or using other backbones, but I am not 100% sure these methods make it outperform the original one. As now the public scores except the final two are hidden. I can't remember which version of DETR achieved the best score. If you are interested, you can view my inference script at all versions here[https://www.kaggle.com/qingyuanwang/detr-inference]

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "976700": "There's a huge change in the leaderboard and it seems that submissions have been checked.\nIt must have taken a lot of effort.Thanks so much to the Kaggle and University of Saskatchewan teams.\n\nNo matter what the final private leaderboard is, we will be very satisfied because we have learned a lot from this competition:).Expecially thanks to the people who share any idea in the forum!\n\nOur team has trained three model:faster-rcnn,EfficientDet,DETR. And we used the fine-tune model to fix the training data.The faster-rcnn+EfficientDet merge score is 0.7649.\nWe will also share our detail solution soon.\n**Btw, will the result be rerun again?**", "976984": "Thanks for sharing @li325040229 \n\nCould you explain DETR more please?\nI also tried to train DETR model but lb is not good.", "977056": "Hi, I am his teammate who worked on DETR. I remember the single model of DETR(no PL, simple TTA WBF) almost (but not) passed 0.7 on the former public set. What about yours?", "977061": "Thanks for reply.\nI commented [this](https://www.kaggle.com/c/global-wheat-detection/discussion/167092).\nWhen i got this scores, i decided to choose good baseline model(e.g. EfficientDet, FasterRCNN, etc.).", "977134": "My training script was mainly forked from [https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr]\nIndeed DETR doesn't beat EfficientDet or FasterRCNN which can easily achieve 0.74+. It is still a feasible method. I will test it more in other contests. I tried adding FPN or using other backbones, but I am not 100% sure these methods make it outperform the original one. As now the public scores except the final two are hidden. I can't remember which version of DETR achieved the best score. If you are interested, you can view my inference script at all versions here[https://www.kaggle.com/qingyuanwang/detr-inference]" }, "source": "meta" }