In this post you will find the most important discoveries published by competition participants. Thank you for sharing and contributing to this competition. These will be discoveries that increase the LB rating. Hope you enjoy this thread.
\n23.11.2021 (no score) EDA - @diegoalejogm is showing us great starfish video created from COTS dataset https://www.kaggle.com/diegoalejogm/great-barrier-reefs-eda-with-animations I am amazed by videos created by @bamps53 I watched all videos to understand how training dataset looks like: https://www.kaggle.com/bamps53/create-annotated-video/notebook
\n23.11.2021 (no score) F2 SCORE - @bamps53 provids us competition metric (f2) implementation https://www.kaggle.com/bamps53/competition-metric-implementation. Another way published by @phalanx you can find in discussion: https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757
\n23.11.2021 (no score) DATASET - @julian3833 in Reef - CV strategy: subsequences! proposed great stratety to divide dataset into folds/splits https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences
\n26.11.2021 (0.453) MODEL - @awsaf49 published Yolo5 infer https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer and training https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer notebooks. As people said here https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757 Yolo5 is one of the best model in this competition so far but we do not see how to achieve score better then 0.453.
\n28.11.2021 (no score) DATA - @soumya9977 presents how to deal with underwater photos and generate new dataset using different techniques https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda
\n30.11.2021 (0.507) MODEL - @remekkinas introduce full YoloX pipeline on Kaggle and … shows how to train model above 0.5 in this competition. Training https://www.kaggle.com/remekkinas/yolox-training-pipeline-cots-dataset-lb-0-507 and inference https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507 notebook. As we can see now … the best LB score using YoloX notebooks is 0.531 now https://www.kaggle.com/ammarnassanalhajali/barrier-reef-yolox-inference
\n23.12.2021 (0.539) MODEL - @parapapapam - https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539 improves YoloX score using object tracking method (Tracker) -> 0.539.
\nWhat discoveries will bring us incoming days? We will see … I will update this topic regullary.
\nUpdated 25.12.2021 - no improvements, no new ideas - best public score 0.539
\nUpdated 26.12.2021 - no improvements, no new ideas - best public score 0.539
27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score. https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/297575. Still the best public score is 0.539.
\nUpdated 28.12.2021 - no improvements, no new ideas … but … more question about Yolo5 and its high score …. maybe somebody will share something interesting … who knows 😄- best public score 0.539
\nUpdated 03.01.2022
\nUpdated 08.01.2022
\nUpdated 13.01.2022
\nIn this post you will find the most important discoveries published by competition participants. Thank you for sharing and contributing to this competition. These will be discoveries that increase the LB rating. Hope you enjoy this thread.
\n23.11.2021 (no score) EDA - @diegoalejogm is showing us great starfish video created from COTS dataset https://www.kaggle.com/diegoalejogm/great-barrier-reefs-eda-with-animations I am amazed by videos created by @bamps53 I watched all videos to understand how training dataset looks like: https://www.kaggle.com/bamps53/create-annotated-video/notebook
\n23.11.2021 (no score) F2 SCORE - @bamps53 provids us competition metric (f2) implementation https://www.kaggle.com/bamps53/competition-metric-implementation. Another way published by @phalanx you can find in discussion: https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757
\n23.11.2021 (no score) DATASET - @julian3833 in Reef - CV strategy: subsequences! proposed great stratety to divide dataset into folds/splits https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences
\n26.11.2021 (0.453) MODEL - @awsaf49 published Yolo5 infer https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer and training https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer notebooks. As people said here https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757 Yolo5 is one of the best model in this competition so far but we do not see how to achieve score better then 0.453.
\n28.11.2021 (no score) DATA - @soumya9977 presents how to deal with underwater photos and generate new dataset using different techniques https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda
\n30.11.2021 (0.507) MODEL - @remekkinas introduce full YoloX pipeline on Kaggle and … shows how to train model above 0.5 in this competition. Training https://www.kaggle.com/remekkinas/yolox-training-pipeline-cots-dataset-lb-0-507 and inference https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507 notebook. As we can see now … the best LB score using YoloX notebooks is 0.531 now https://www.kaggle.com/ammarnassanalhajali/barrier-reef-yolox-inference
\n23.12.2021 (0.539) MODEL - @parapapapam - https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539 improves YoloX score using object tracking method (Tracker) -> 0.539.
\nWhat discoveries will bring us incoming days? We will see … I will update this topic regullary.
\nUpdated 25.12.2021 - no improvements, no new ideas - best public score 0.539
\nUpdated 26.12.2021 - no improvements, no new ideas - best public score 0.539
27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score. https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/297575. Still the best public score is 0.539.
\nUpdated 28.12.2021 - no improvements, no new ideas … but … more question about Yolo5 and its high score …. maybe somebody will share something interesting … who knows 😄- best public score 0.539
\nUpdated 03.01.2022
\nUpdated 08.01.2022
\nUpdated 13.01.2022
\nNice post 👍
", "rawMarkdown": "Nice post 👍", "votes": null }, { "id": "1630016", "postDate": "12/26/2021 21:50:52", "content": "Thank you! This is for all who wants share and learn :)
", "rawMarkdown": "Thank you! This is for all who wants share and learn :)", "votes": null }, { "id": "1630918", "postDate": "12/27/2021 20:53:29", "content": "Nice post, thank you for sharing
", "rawMarkdown": "Nice post, thank you for sharing", "votes": null }, { "id": "1630921", "postDate": "12/27/2021 21:04:22", "content": "You are welcome. I will update it :)
", "rawMarkdown": "You are welcome. I will update it :)", "votes": null }, { "id": "1630928", "postDate": "12/27/2021 21:16:10", "content": "\n\n27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score.
\n
How is this an upgrade if YoloX score is higher 5.39?
", "rawMarkdown": "> 27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score.\n\nHow is this an upgrade if YoloX score is higher 5.39?", "votes": null }, { "id": "1630936", "postDate": "12/27/2021 21:21:35", "content": "Yes. I agree. As I can see in this notebook is not progress but … I am just sharing idea … wbf.
", "rawMarkdown": "Yes. I agree. As I can see in this notebook is not progress but ... I am just sharing idea ... wbf.", "votes": null }, { "id": "1631141", "postDate": "12/28/2021 04:48:05", "content": "Hi! I achieved 0.491 using Yolo5 Great-Barrier-Reef: YOLOv5 [infer] 🌊.
\nI change some parameters but when i set the mosaic_prob from 1 to 0.5 it got better, reading the Yolo5 GitHub they don't recommend using it on a small dataset and it is also recommended for datasets like COCO (many classes, scales, etc), so on this competition in my own opinion it should be decreased.
\nTraining notebook: OWN Great-Barrier-Reef: YOLOv5 [train] 🌊\n
Thank you for sharing ideas! Great! 👍
", "rawMarkdown": "Thank you for sharing ideas! Great! 👍", "votes": null }, { "id": "1631223", "postDate": "12/28/2021 07:28:00", "content": "Thanks! I'm now working with YoloX, quick question, your .539 model was trained on a pre-trained cots model?
", "rawMarkdown": "Thanks! I'm now working with YoloX, quick question, your .539 model was trained on a pre-trained cots model?", "votes": null }, { "id": "1631231", "postDate": "12/28/2021 07:35:20", "content": "No. All models I have trained (even 0.55x) was on standard YoloX checkpoint. I am still looking for one more tip which improves score significantly (let say jump to 0.58-0.59). Now it is … small increase I think compare to guys from TOP who talk about yolo5 and 0.59 without no problem :)
", "rawMarkdown": "No. All models I have trained (even 0.55x) was on standard YoloX checkpoint. I am still looking for one more tip which improves score significantly (let say jump to 0.58-0.59). Now it is ... small increase I think compare to guys from TOP who talk about yolo5 and 0.59 without no problem :)", "votes": null }, { "id": "1631522", "postDate": "12/28/2021 13:50:44", "content": "a quick question, where did you find all those config-parameter for YOLOX ? I just found very little on Github and was really surprised when I found self.act = 'silu' can be set in config
", "rawMarkdown": "a quick question, where did you find all those config-parameter for YOLOX ? I just found very little on Github and was really surprised when I found self.act = 'silu' can be set in config", "votes": null }, { "id": "1631537", "postDate": "12/28/2021 14:18:43", "content": "Exp class :) https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/exp/yolox_base.py
", "rawMarkdown": "Exp class :) https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/exp/yolox_base.py", "votes": null }, { "id": "1632243", "postDate": "12/29/2021 12:17:27", "content": "Thanks for Sharing ! It was Awesome work!
", "rawMarkdown": "Thanks for Sharing ! It was Awesome work!", "votes": null }, { "id": "1632269", "postDate": "12/29/2021 12:43:37", "content": "Thanks soooo much!
", "rawMarkdown": "Thanks soooo much!", "votes": null }, { "id": "1632273", "postDate": "12/29/2021 12:44:35", "content": "Thank you very much. It will be updated regularly:)
", "rawMarkdown": "Thank you very much. It will be updated regularly:)", "votes": null }, { "id": "1632327", "postDate": "12/29/2021 13:56:01", "content": "I bet I can achieve a score better than 0.453 using YoloX notebooks. What is preventing me is how I'm being treated like a novice. This attitude is not bringing the best out of me. I hope this changes.
", "rawMarkdown": "I bet I can achieve a score better than 0.453 using YoloX notebooks. What is preventing me is how I'm being treated like a novice. This attitude is not bringing the best out of me. I hope this changes.", "votes": null }, { "id": "1632366", "postDate": "12/29/2021 14:41:48", "content": "One way you can improve is experimenting. I am sure you improve your score soon. See me on LB … how many our team submitted… and we have big problem to improve … but … this is sign for us … think and make more experiments:)
", "rawMarkdown": "One way you can improve is experimenting. I am sure you improve your score soon. See me on LB … how many our team submitted… and we have big problem to improve … but … this is sign for us … think and make more experiments:)", "votes": null }, { "id": "1633061", "postDate": "12/30/2021 11:21:51", "content": "What improvements have you made to jump to TOP?
", "rawMarkdown": "What improvements have you made to jump to TOP?", "votes": null }, { "id": "1633403", "postDate": "12/30/2021 18:25:01", "content": "We use everything we have shared in this competition. Look here:
\n@csarolivares Can you say which yolov5 was used to achieve 0.491? s, m, l or x?
\nGreat job!
Using yolov5m
", "rawMarkdown": "Using yolov5m", "votes": null }, { "id": "1638910", "postDate": "01/05/2022 06:45:52", "content": "Hi, thank you for a nice overview. As the evaluation metric is computed for various IOU thresholds, it seems reasonable to have precise detections. We worked on a general refinement network; maybe, it can be helpful in this competition, see details in https://graphicwg.irafm.osu.cz/storage/bbrefinement.pdf
", "rawMarkdown": "Hi, thank you for a nice overview. As the evaluation metric is computed for various IOU thresholds, it seems reasonable to have precise detections. We worked on a general refinement network; maybe, it can be helpful in this competition, see details in https://graphicwg.irafm.osu.cz/storage/bbrefinement.pdf", "votes": null }, { "id": "1648746", "postDate": "01/13/2022 16:54:13", "content": "Quick question. Has anyone used WBF and had success with it? I have seen a few kernels and tried it on my own but it always seems to produce worse results. Any tips or similar experiences?
", "rawMarkdown": "Quick question. Has anyone used WBF and had success with it? I have seen a few kernels and tried it on my own but it always seems to produce worse results. Any tips or similar experiences?", "votes": null }, { "id": "1665458", "postDate": "01/26/2022 21:24:18", "content": "@remekkinas Really appreciate your contribution (in terms of code, discussions and comments) in this competition. It is great to see someone high up on the LB helping the community with the resources… Really helpful for me as a beginner…..Thanks
", "rawMarkdown": "remekkinas Really appreciate your contribution (in terms of code, discussions and comments) in this competition. It is great to see someone high up on the LB helping the community with the resources... Really helpful for me as a beginner.....Thanks", "votes": null }, { "id": "1665461", "postDate": "01/26/2022 21:26:07", "content": "@venkateshkulkarni11 thank you very much …. so it is possible to be high on LB and share 😄😆😍
", "rawMarkdown": "venkateshkulkarni11 thank you very much .... so it is possible to be high on LB and share 😄😆😍", "votes": null }, { "id": "1669705", "postDate": "01/30/2022 21:44:06", "content": "HaHa, doing many trials but gain not improvement is so exhausting. But you got such a great PB, man! It is inspiring!
", "rawMarkdown": "HaHa, doing many trials but gain not improvement is so exhausting. But you got such a great PB, man! It is inspiring!", "votes": null } ], "comments": [ { "id": 1630011, "author_name": "arelahi", "author_url": "", "post_date": "12/26/2021 21:43:10", "content": "Nice post 👍
", "votes": null, "replies": [ { "id": 1630016, "author_name": "remekkinas", "author_url": "", "post_date": "12/26/2021 21:50:52", "content": "Thank you! This is for all who wants share and learn :)
", "votes": null, "replies": [] } ] }, { "id": 1630918, "author_name": "deryasolak", "author_url": "", "post_date": "12/27/2021 20:53:29", "content": "Nice post, thank you for sharing
", "votes": null, "replies": [ { "id": 1630921, "author_name": "remekkinas", "author_url": "", "post_date": "12/27/2021 21:04:22", "content": "You are welcome. I will update it :)
", "votes": null, "replies": [] } ] }, { "id": 1630928, "author_name": "diegoalejogm", "author_url": "", "post_date": "12/27/2021 21:16:10", "content": "\n\n27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score.
\n
How is this an upgrade if YoloX score is higher 5.39?
", "votes": null, "replies": [ { "id": 1630936, "author_name": "remekkinas", "author_url": "", "post_date": "12/27/2021 21:21:35", "content": "Yes. I agree. As I can see in this notebook is not progress but … I am just sharing idea … wbf.
", "votes": null, "replies": [] } ] }, { "id": 1631141, "author_name": "csarolivares", "author_url": "", "post_date": "12/28/2021 04:48:05", "content": "Hi! I achieved 0.491 using Yolo5 Great-Barrier-Reef: YOLOv5 [infer] 🌊.
\nI change some parameters but when i set the mosaic_prob from 1 to 0.5 it got better, reading the Yolo5 GitHub they don't recommend using it on a small dataset and it is also recommended for datasets like COCO (many classes, scales, etc), so on this competition in my own opinion it should be decreased.
\nTraining notebook: OWN Great-Barrier-Reef: YOLOv5 [train] 🌊\n
Thank you for sharing ideas! Great! 👍
", "votes": null, "replies": [] }, { "id": 1631223, "author_name": "csarolivares", "author_url": "", "post_date": "12/28/2021 07:28:00", "content": "Thanks! I'm now working with YoloX, quick question, your .539 model was trained on a pre-trained cots model?
", "votes": null, "replies": [] }, { "id": 1631231, "author_name": "remekkinas", "author_url": "", "post_date": "12/28/2021 07:35:20", "content": "No. All models I have trained (even 0.55x) was on standard YoloX checkpoint. I am still looking for one more tip which improves score significantly (let say jump to 0.58-0.59). Now it is … small increase I think compare to guys from TOP who talk about yolo5 and 0.59 without no problem :)
", "votes": null, "replies": [] }, { "id": 1633061, "author_name": "willor", "author_url": "", "post_date": "12/30/2021 11:21:51", "content": "What improvements have you made to jump to TOP?
", "votes": null, "replies": [] }, { "id": 1633403, "author_name": "remekkinas", "author_url": "", "post_date": "12/30/2021 18:25:01", "content": "We use everything we have shared in this competition. Look here:
\n@csarolivares Can you say which yolov5 was used to achieve 0.491? s, m, l or x?
\nGreat job!
Using yolov5m
", "votes": null, "replies": [] } ] }, { "id": 1631522, "author_name": "andylaer", "author_url": "", "post_date": "12/28/2021 13:50:44", "content": "a quick question, where did you find all those config-parameter for YOLOX ? I just found very little on Github and was really surprised when I found self.act = 'silu' can be set in config
", "votes": null, "replies": [ { "id": 1631537, "author_name": "remekkinas", "author_url": "", "post_date": "12/28/2021 14:18:43", "content": "Exp class :) https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/exp/yolox_base.py
", "votes": null, "replies": [] }, { "id": 1632269, "author_name": "andylaer", "author_url": "", "post_date": "12/29/2021 12:43:37", "content": "Thanks soooo much!
", "votes": null, "replies": [] } ] }, { "id": 1632243, "author_name": "balavashan", "author_url": "", "post_date": "12/29/2021 12:17:27", "content": "Thanks for Sharing ! It was Awesome work!
", "votes": null, "replies": [ { "id": 1632273, "author_name": "remekkinas", "author_url": "", "post_date": "12/29/2021 12:44:35", "content": "Thank you very much. It will be updated regularly:)
", "votes": null, "replies": [] } ] }, { "id": 1632327, "author_name": "adeyemiadewole", "author_url": "", "post_date": "12/29/2021 13:56:01", "content": "I bet I can achieve a score better than 0.453 using YoloX notebooks. What is preventing me is how I'm being treated like a novice. This attitude is not bringing the best out of me. I hope this changes.
", "votes": null, "replies": [ { "id": 1632366, "author_name": "remekkinas", "author_url": "", "post_date": "12/29/2021 14:41:48", "content": "One way you can improve is experimenting. I am sure you improve your score soon. See me on LB … how many our team submitted… and we have big problem to improve … but … this is sign for us … think and make more experiments:)
", "votes": null, "replies": [] } ] }, { "id": 1638910, "author_name": "petrhurtik", "author_url": "", "post_date": "01/05/2022 06:45:52", "content": "Hi, thank you for a nice overview. As the evaluation metric is computed for various IOU thresholds, it seems reasonable to have precise detections. We worked on a general refinement network; maybe, it can be helpful in this competition, see details in https://graphicwg.irafm.osu.cz/storage/bbrefinement.pdf
", "votes": null, "replies": [] }, { "id": 1648746, "author_name": "gtownfoster", "author_url": "", "post_date": "01/13/2022 16:54:13", "content": "Quick question. Has anyone used WBF and had success with it? I have seen a few kernels and tried it on my own but it always seems to produce worse results. Any tips or similar experiences?
", "votes": null, "replies": [] }, { "id": 1665458, "author_name": "venkateshkulkarni11", "author_url": "", "post_date": "01/26/2022 21:24:18", "content": "@remekkinas Really appreciate your contribution (in terms of code, discussions and comments) in this competition. It is great to see someone high up on the LB helping the community with the resources… Really helpful for me as a beginner…..Thanks
", "votes": null, "replies": [ { "id": 1665461, "author_name": "remekkinas", "author_url": "", "post_date": "01/26/2022 21:26:07", "content": "@venkateshkulkarni11 thank you very much …. so it is possible to be high on LB and share 😄😆😍
", "votes": null, "replies": [] } ] }, { "id": 1669705, "author_name": "calvchen", "author_url": "", "post_date": "01/30/2022 21:44:06", "content": "HaHa, doing many trials but gain not improvement is so exhausting. But you got such a great PB, man! It is inspiring!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1626846": "In this post you will find the most important discoveries published by competition participants. Thank you for sharing and contributing to this competition. These will be discoveries that increase the LB rating. Hope you enjoy this thread. \n\n23.11.2021 (no score) **EDA** - @diegoalejogm is showing us great starfish video created from COTS dataset https://www.kaggle.com/diegoalejogm/great-barrier-reefs-eda-with-animations I am amazed by videos created by @bamps53 I watched all videos to understand how training dataset looks like: https://www.kaggle.com/bamps53/create-annotated-video/notebook\n\n23.11.2021 (no score) **F2 SCORE** - @bamps53 provids us competition metric (f2) implementation https://www.kaggle.com/bamps53/competition-metric-implementation. Another way published by @phalanx you can find in discussion: https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757 \n\n23.11.2021 (no score) **DATASET** - @julian3833 in Reef - CV strategy: subsequences! proposed great stratety to divide dataset into folds/splits https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences\n\n26.11.2021 (0.453) **MODEL** - @awsaf49 published Yolo5 infer https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer and training https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-infer notebooks. As people said here https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757 Yolo5 is one of the best model in this competition so far but we do not see how to achieve score better then 0.453.\n\n28.11.2021 (no score) **DATA** - @soumya9977 presents how to deal with underwater photos and generate new dataset using different techniques https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda\n\n30.11.2021 (0.507) **MODEL** - @remekkinas introduce full YoloX pipeline on Kaggle and ... shows how to train model above 0.5 in this competition. Training https://www.kaggle.com/remekkinas/yolox-training-pipeline-cots-dataset-lb-0-507 and inference https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507 notebook. As we can see now ... the best LB score using YoloX notebooks is 0.531 now https://www.kaggle.com/ammarnassanalhajali/barrier-reef-yolox-inference\n\n23.12.2021 (0.539) **MODEL** - @parapapapam - https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539 improves YoloX score using object tracking method (Tracker) -> 0.539. \n\nWhat discoveries will bring us incoming days? We will see ... I will update this topic regullary. \n\n**Updated 25.12.2021** - no improvements, no new ideas - best public score 0.539\n**Updated 26.12.2021** - no improvements, no new ideas - best public score 0.539\n\n27.12.2021 **MODEL** - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score. https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/297575. Still the best public score is 0.539.\n\n**Updated 28.12.2021** - no improvements, no new ideas ... but ... more question about Yolo5 and its high score .... maybe somebody will share something interesting ... who knows 😄- best public score 0.539\n\n**Updated 03.01.2022**\n- Great post about augumentations in CV by @andradaolteanu https://www.kaggle.com/andradaolteanu/greatbarrierreef-full-guide-to-bboxaugmentation\n\n**Updated 08.01.2022**\n- no changes\n\n**Updated 13.01.2022**\n- interesting notebook Yolov5 is all you need https://www.kaggle.com/steamedsheep/yolov5-is-all-you-need by @steamedsheep - best public score so far 0.579 which is great improvement from previous one.", "1630011": "Nice post 👍", "1630016": "Thank you! This is for all who wants share and learn :)", "1630918": "Nice post, thank you for sharing", "1630921": "You are welcome. I will update it :)", "1630928": "> 27.12.2021 MODEL - @yamqwe shows how to jump to 0.522 using wbf (Weighted Boxes Fusion) using two models Yolov5 and YoloX. It is definetely way to improve score.\n\nHow is this an upgrade if YoloX score is higher 5.39?", "1630936": "Yes. I agree. As I can see in this notebook is not progress but ... I am just sharing idea ... wbf.", "1631141": "Hi! I achieved 0.491 using Yolo5 [Great-Barrier-Reef: YOLOv5 [infer] 🌊](https://www.kaggle.com/csarolivares/great-barrier-reef-yolov5-infer).\nI change some parameters but when i set the mosaic_prob from 1 to 0.5 it got better, reading the Yolo5 GitHub they don't recommend using it on a small dataset and it is also recommended for datasets like COCO (many classes, scales, etc), so on this competition **in my own opinion** it should be decreased.\nTraining notebook: [OWN Great-Barrier-Reef: YOLOv5 [train] 🌊\n](https://www.kaggle.com/csarolivares/own-great-barrier-reef-yolov5-train)", "1631219": "Thank you for sharing ideas! Great! 👍", "1631223": "Thanks! I'm now working with YoloX, quick question, your .539 model was trained on a pre-trained cots model?", "1631231": "No. All models I have trained (even 0.55x) was on standard YoloX checkpoint. I am still looking for one more tip which improves score significantly (let say jump to 0.58-0.59). Now it is ... small increase I think compare to guys from TOP who talk about yolo5 and 0.59 without no problem :)", "1631522": "a quick question, where did you find all those config-parameter for YOLOX ? I just found very little on Github and was really surprised when I found self.act = 'silu' can be set in config", "1631537": "Exp class :) https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/exp/yolox_base.py", "1632243": "Thanks for Sharing ! It was Awesome work!", "1632269": "Thanks soooo much!", "1632273": "Thank you very much. It will be updated regularly:)", "1632327": "I bet I can achieve a score better than 0.453 using YoloX notebooks. What is preventing me is how I'm being treated like a novice. This attitude is not bringing the best out of me. I hope this changes.", "1632366": "One way you can improve is experimenting. I am sure you improve your score soon. See me on LB … how many our team submitted… and we have big problem to improve … but … this is sign for us … think and make more experiments:)", "1633061": "What improvements have you made to jump to TOP?", "1633403": "We use everything we have shared in this competition. Look here:\n- https://www.kaggle.com/remekkinas/yolor-p6-w6-one-more-yolo-on-kaggle-train\n- https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507\n- https://www.kaggle.com/remekkinas/yolox-training-pipeline-cots-dataset-lb-0-507", "1635601": "csarolivares Can you say which yolov5 was used to achieve 0.491? s, m, l or x? \nGreat job!", "1635627": "Using yolov5m", "1638910": "Hi, thank you for a nice overview. As the evaluation metric is computed for various IOU thresholds, it seems reasonable to have precise detections. We worked on a general refinement network; maybe, it can be helpful in this competition, see details in https://graphicwg.irafm.osu.cz/storage/bbrefinement.pdf", "1648746": "Quick question. Has anyone used WBF and had success with it? I have seen a few kernels and tried it on my own but it always seems to produce worse results. Any tips or similar experiences?", "1665458": "remekkinas Really appreciate your contribution (in terms of code, discussions and comments) in this competition. It is great to see someone high up on the LB helping the community with the resources... Really helpful for me as a beginner.....Thanks", "1665461": "venkateshkulkarni11 thank you very much .... so it is possible to be high on LB and share 😄😆😍", "1669705": "HaHa, doing many trials but gain not improvement is so exhausting. But you got such a great PB, man! It is inspiring!" }, "source": "meta" }