![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F216357%2F9379c355271224374212ea563e5b7a8e%2FClipboard_02-28-2026_01.jpg?generation=1772247981493470&alt=media\")
Thanks to organisers for a fun ride and congrats to the winners!
\nThe very complicated metric implied that postprocessing would be playing a huge role. My solution was quite simple - a geo-mean blend of two 3D LowRes models (different patch size, normalization, etc) followed by 3D Cascade Fullres. The latter takes care of the postprocessing quite well. Even simple thresholding would brings to a top20-top30 region. A 5-step public post-processing on top covers the last mile. In fact, changing radiuses from 3-2 to 2-1 would have brought me to the 3rd place - but it was not giving any visual benefits on validation.
Overall very happy with the 🥇11th place given my recent painfull experience of missing the Gold train by 1 position. Twice. In a row 😁
\nI did not quite go far enough to use AI to write the complete solution but it was indeed indispensible to explain how nnUnetV2 works, where to change the optimiser from sgd to adam (which saves a lot in training compute btw), where to change tta, step size, things like that.
\nCheers!
", "messages": [ { "id": "3414988", "postDate": "02/28/2026 03:03:14", "content": "Thanks to organisers for a fun ride and congrats to the winners!
\nThe very complicated metric implied that postprocessing would be playing a huge role. My solution was quite simple - a geo-mean blend of two 3D LowRes models (different patch size, normalization, etc) followed by 3D Cascade Fullres. The latter takes care of the postprocessing quite well. Even simple thresholding would brings to a top20-top30 region. A 5-step public post-processing on top covers the last mile. In fact, changing radiuses from 3-2 to 2-1 would have brought me to the 3rd place - but it was not giving any visual benefits on validation.
Overall very happy with the 🥇11th place given my recent painfull experience of missing the Gold train by 1 position. Twice. In a row 😁
\nI did not quite go far enough to use AI to write the complete solution but it was indeed indispensible to explain how nnUnetV2 works, where to change the optimiser from sgd to adam (which saves a lot in training compute btw), where to change tta, step size, things like that.
\nCheers!
", "rawMarkdown": "Thanks to organisers for a fun ride and congrats to the winners!\n\nThe very complicated metric implied that postprocessing would be playing a huge role. My solution was quite simple - a geo-mean blend of two 3D LowRes models (different patch size, normalization, etc) followed by 3D Cascade Fullres. The latter takes care of the postprocessing quite well. Even simple thresholding would brings to a top20-top30 region. A 5-step public post-processing on top covers the last mile. In fact, changing radiuses from 3-2 to 2-1 would have brought me to the 3rd place - but it was not giving any visual benefits on validation. \n\nOverall **very happy** with the 🥇11th place given my recent painfull experience of missing the Gold train by 1 position. Twice. In a row 😁\n\nI did not quite go far enough to use AI to write the complete solution but it was indeed indispensible to explain how nnUnetV2 works, where to change the optimiser from sgd to adam (which saves a lot in training compute btw), where to change tta, step size, things like that. \n\nCheers!", "votes": null }, { "id": "3415088", "postDate": "02/28/2026 06:40:41", "content": "Nice work. I think you would be uprank to price zone. Curretnly the 6th place looks suspicious. Probably someone from 7th place scam the place. They got similar score. Not to mention, these 6th place guys are so bright and intelligent that, they just joined in kaggle in less than a month and got top place.
\n", "rawMarkdown": "Nice work. I think you would be uprank to price zone. Curretnly the 6th place looks suspicious. Probably someone from 7th place scam the place. They got similar score. Not to mention, these 6th place guys are so bright and intelligent that, they just joined in kaggle in less than a month and got top place. \n\ncc @giorgioangelotti @sohier", "votes": null }, { "id": "3415218", "postDate": "02/28/2026 12:42:15", "content": "If you have specific evidence they cheated, then share it with compliance@kaggle. Public finger pointing without evidence is despicable.
\nThat some people do well at their first competition is not unusual, especially from China. They have some kaggle like platforms and they train on it before trying Kaggle.
\nMoreover, if people have been working on similar topic in their work (3D segmentation), then they may just do well. The person from nnunet team was in gold at his first submission before data reset. Did you found it suspicious? I didn't.
", "rawMarkdown": "If you have specific evidence they cheated, then share it with compliance@kaggle. Public finger pointing without evidence is despicable.\n\nThat some people do well at their first competition is not unusual, especially from China. They have some kaggle like platforms and they train on it before trying Kaggle.\n\nMoreover, if people have been working on similar topic in their work (3D segmentation), then they may just do well. The person from nnunet team was in gold at his first submission before data reset. Did you found it suspicious? I didn't.", "votes": null }, { "id": "3415995", "postDate": "03/01/2026 19:09:58", "content": "Okay, I will wait for their reproducible write up solution. Its better not to be the same solution as team 7.
", "rawMarkdown": "Okay, I will wait for their reproducible write up solution. Its better not to be the same solution as team 7.", "votes": null } ], "comments": [ { "id": 3415088, "author_name": "jannatul21", "author_url": "", "post_date": "02/28/2026 06:40:41", "content": "Nice work. I think you would be uprank to price zone. Curretnly the 6th place looks suspicious. Probably someone from 7th place scam the place. They got similar score. Not to mention, these 6th place guys are so bright and intelligent that, they just joined in kaggle in less than a month and got top place.
\n", "votes": null, "replies": [ { "id": 3415218, "author_name": "cpmpml", "author_url": "", "post_date": "02/28/2026 12:42:15", "content": "If you have specific evidence they cheated, then share it with compliance@kaggle. Public finger pointing without evidence is despicable.
\nThat some people do well at their first competition is not unusual, especially from China. They have some kaggle like platforms and they train on it before trying Kaggle.
\nMoreover, if people have been working on similar topic in their work (3D segmentation), then they may just do well. The person from nnunet team was in gold at his first submission before data reset. Did you found it suspicious? I didn't.
", "votes": null, "replies": [ { "id": 3415995, "author_name": "jannatul21", "author_url": "", "post_date": "03/01/2026 19:09:58", "content": "Okay, I will wait for their reproducible write up solution. Its better not to be the same solution as team 7.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3414988": "Thanks to organisers for a fun ride and congrats to the winners!\n\nThe very complicated metric implied that postprocessing would be playing a huge role. My solution was quite simple - a geo-mean blend of two 3D LowRes models (different patch size, normalization, etc) followed by 3D Cascade Fullres. The latter takes care of the postprocessing quite well. Even simple thresholding would brings to a top20-top30 region. A 5-step public post-processing on top covers the last mile. In fact, changing radiuses from 3-2 to 2-1 would have brought me to the 3rd place - but it was not giving any visual benefits on validation. \n\nOverall **very happy** with the 🥇11th place given my recent painfull experience of missing the Gold train by 1 position. Twice. In a row 😁\n\nI did not quite go far enough to use AI to write the complete solution but it was indeed indispensible to explain how nnUnetV2 works, where to change the optimiser from sgd to adam (which saves a lot in training compute btw), where to change tta, step size, things like that. \n\nCheers!", "3415088": "Nice work. I think you would be uprank to price zone. Curretnly the 6th place looks suspicious. Probably someone from 7th place scam the place. They got similar score. Not to mention, these 6th place guys are so bright and intelligent that, they just joined in kaggle in less than a month and got top place. \n\ncc @giorgioangelotti @sohier", "3415218": "If you have specific evidence they cheated, then share it with compliance@kaggle. Public finger pointing without evidence is despicable.\n\nThat some people do well at their first competition is not unusual, especially from China. They have some kaggle like platforms and they train on it before trying Kaggle.\n\nMoreover, if people have been working on similar topic in their work (3D segmentation), then they may just do well. The person from nnunet team was in gold at his first submission before data reset. Did you found it suspicious? I didn't.", "3415995": "Okay, I will wait for their reproducible write up solution. Its better not to be the same solution as team 7." }, "source": "meta" }