For me the it's quite long.....
\n\nAlso I suggest that the daily submission limit should be increased to 10-20 for a unstable dataset like this. In a way, public LB should be allowed to probed so that we can gauge the difference in distribution between test and train.
\n\nYou're welcomed to add more!
", "messages": [ { "id": "958269", "postDate": "08/04/2020 22:21:34", "content": "For me the it's quite long.....
\n\nAlso I suggest that the daily submission limit should be increased to 10-20 for a unstable dataset like this. In a way, public LB should be allowed to probed so that we can gauge the difference in distribution between test and train.
\n\nYou're welcomed to add more!
", "rawMarkdown": "For me the it's quite long.....\n\n- mmdetection (only managed to get to 0.73, but I did not implement mixup/mosaic myself)\n- Generalized Focal Loss\n- Guided Anchoring\n- DetectoRS/recursive feature pyramid\n- Integral Loss (I tried to design a cascade architecture as well but it was too hard to optimize)\n- Messing up with anchors/target assigners\n- DIoU/CIoU Loss\n- Reppoints V2\n- FCOS\n- ATSS\n- Stitcher Augmentation\n- DETR, or generally using a transformer as a feature encoder (I designed two models myself, none worked)\n- FixMatch's detection variant (Also self-designed. I will try to make it work after the entire dataset is released and maybe write a tech report/paper)\n- Autoaugment a la carte, meaning that I didn't search for the augmentations myself\n- Consistency regularization based on feature maps in general\n- SPIKE Dataset (I didn't use their labels; I re-pseudo-labelled the dataset myself)\n- AdaBN\n- NOH-NMS/Adaptive NMS\n- Cluster NMS (takes too long and wasn't significantly different from batched nms)\n- AdjNMS (won OID last year, but gives horrible results in this competition for some reason)\n- Anything that is not COCO-pretrained entirely\n- Exponential Weight Averaging\n- CycleLR \n- Free Anchor\n- GHM/GHC Loss\n\nAlso I suggest that the daily submission limit should be increased to 10-20 for a unstable dataset like this. In a way, public LB should be allowed to probed so that we can gauge the difference in distribution between test and train.\n\nYou're welcomed to add more!", "votes": null }, { "id": "958286", "postDate": "08/04/2020 22:56:12", "content": "YoloV5
", "rawMarkdown": "YoloV5", "votes": null }, { "id": "958290", "postDate": "08/04/2020 23:00:40", "content": "Thanks, that's an impressive list. Also interested in what did work 😉
", "rawMarkdown": "Thanks, that's an impressive list. Also interested in what did work 😉", "votes": null }, { "id": "958303", "postDate": "08/04/2020 23:30:34", "content": "For me, I was very very disappointed with yolov4. Reading some medium articles and such, yolov4 is supposedly both faster and more accurate than yolov5, but with WBF/TTA, I couldn't get higher than about 0.715 and with PL/OOF, 0.735. For how much time I put in creating the training pipeline as well as figuring out how to run Darknet (I had so much trouble running Darknet on Kaggle until someone helped me out) to get the licence compliant, I'm quite disappointed. Should've gone with FRCNN I guess. Great list, most of which I didn't even think about.
", "rawMarkdown": "For me, I was very very disappointed with yolov4. Reading some medium articles and such, yolov4 is supposedly both faster and more accurate than yolov5, but with WBF/TTA, I couldn't get higher than about 0.715 and with PL/OOF, 0.735. For how much time I put in creating the training pipeline as well as figuring out how to run Darknet (I had so much trouble running Darknet on Kaggle until someone helped me out) to get the licence compliant, I'm quite disappointed. Should've gone with FRCNN I guess. Great list, most of which I didn't even think about.", "votes": null }, { "id": "958304", "postDate": "08/04/2020 23:32:42", "content": "I didn't do much of what you mentioned. \n* YoloV3, V4\n* DetectoRS\n* LabelSmoothing\n* etc...
\n\nI look forward to your solution.😊
", "rawMarkdown": "I didn't do much of what you mentioned. \n* YoloV3, V4\n* DetectoRS\n* LabelSmoothing\n* etc...\n\nI look forward to your solution.😊", "votes": null }, { "id": "958384", "postDate": "08/05/2020 00:40:32", "content": "I am using CenterNet with EfficientNet b5 + BiFPN backbone. For me it's super short haha
\n\nAfter the shake? Nothing :(
", "rawMarkdown": "After the shake? Nothing :(", "votes": null }, { "id": "958571", "postDate": "08/05/2020 04:06:34", "content": "", "rawMarkdown": "", "votes": null }, { "id": "958587", "postDate": "08/05/2020 04:22:36", "content": "in the end mmdet take top1)
", "rawMarkdown": "in the end mmdet take top1)", "votes": null } ], "comments": [ { "id": 958286, "author_name": "jonykarki", "author_url": "", "post_date": "08/04/2020 22:56:12", "content": "YoloV5
", "votes": null, "replies": [] }, { "id": 958290, "author_name": "datahobbit", "author_url": "", "post_date": "08/04/2020 23:00:40", "content": "Thanks, that's an impressive list. Also interested in what did work 😉
", "votes": null, "replies": [ { "id": 958409, "author_name": "alexanderliao", "author_url": "", "post_date": "08/05/2020 01:07:25", "content": "After the shake? Nothing :(
", "votes": null, "replies": [] } ] }, { "id": 958303, "author_name": "stanleyjzheng", "author_url": "", "post_date": "08/04/2020 23:30:34", "content": "For me, I was very very disappointed with yolov4. Reading some medium articles and such, yolov4 is supposedly both faster and more accurate than yolov5, but with WBF/TTA, I couldn't get higher than about 0.715 and with PL/OOF, 0.735. For how much time I put in creating the training pipeline as well as figuring out how to run Darknet (I had so much trouble running Darknet on Kaggle until someone helped me out) to get the licence compliant, I'm quite disappointed. Should've gone with FRCNN I guess. Great list, most of which I didn't even think about.
", "votes": null, "replies": [] }, { "id": 958304, "author_name": "piantic", "author_url": "", "post_date": "08/04/2020 23:32:42", "content": "I didn't do much of what you mentioned. \n* YoloV3, V4\n* DetectoRS\n* LabelSmoothing\n* etc...
\n\nI look forward to your solution.😊
", "votes": null, "replies": [] }, { "id": 958384, "author_name": "jihangz", "author_url": "", "post_date": "08/05/2020 00:40:32", "content": "I am using CenterNet with EfficientNet b5 + BiFPN backbone. For me it's super short haha
\n\nin the end mmdet take top1)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "958269": "For me the it's quite long.....\n\n- mmdetection (only managed to get to 0.73, but I did not implement mixup/mosaic myself)\n- Generalized Focal Loss\n- Guided Anchoring\n- DetectoRS/recursive feature pyramid\n- Integral Loss (I tried to design a cascade architecture as well but it was too hard to optimize)\n- Messing up with anchors/target assigners\n- DIoU/CIoU Loss\n- Reppoints V2\n- FCOS\n- ATSS\n- Stitcher Augmentation\n- DETR, or generally using a transformer as a feature encoder (I designed two models myself, none worked)\n- FixMatch's detection variant (Also self-designed. I will try to make it work after the entire dataset is released and maybe write a tech report/paper)\n- Autoaugment a la carte, meaning that I didn't search for the augmentations myself\n- Consistency regularization based on feature maps in general\n- SPIKE Dataset (I didn't use their labels; I re-pseudo-labelled the dataset myself)\n- AdaBN\n- NOH-NMS/Adaptive NMS\n- Cluster NMS (takes too long and wasn't significantly different from batched nms)\n- AdjNMS (won OID last year, but gives horrible results in this competition for some reason)\n- Anything that is not COCO-pretrained entirely\n- Exponential Weight Averaging\n- CycleLR \n- Free Anchor\n- GHM/GHC Loss\n\nAlso I suggest that the daily submission limit should be increased to 10-20 for a unstable dataset like this. In a way, public LB should be allowed to probed so that we can gauge the difference in distribution between test and train.\n\nYou're welcomed to add more!", "958286": "YoloV5", "958290": "Thanks, that's an impressive list. Also interested in what did work 😉", "958303": "For me, I was very very disappointed with yolov4. Reading some medium articles and such, yolov4 is supposedly both faster and more accurate than yolov5, but with WBF/TTA, I couldn't get higher than about 0.715 and with PL/OOF, 0.735. For how much time I put in creating the training pipeline as well as figuring out how to run Darknet (I had so much trouble running Darknet on Kaggle until someone helped me out) to get the licence compliant, I'm quite disappointed. Should've gone with FRCNN I guess. Great list, most of which I didn't even think about.", "958304": "I didn't do much of what you mentioned. \n* YoloV3, V4\n* DetectoRS\n* LabelSmoothing\n* etc...\n\nI look forward to your solution.😊", "958384": "I am using CenterNet with EfficientNet b5 + BiFPN backbone. For me it's super short haha\n\n1. Pseudo label - it gives +0.0002 on private LB, but -0.003 on public, so I didn't select it as my final submission\n2. Backbones larger than b5 - I only tried efficientnet-b6 and it was worse than b5\n3. Ensemble too many folds - for me 3-fold WBF better than 5-fold. Wondering why\n4. Large batch by gradient accumulation\n5. EMA", "958409": "After the shake? Nothing :(", "958571": "", "958587": "in the end mmdet take top1)" }, "source": "meta" }