{ "id": 475126, "title": "did we get tricked? the voi is the key", "url": "/competitions/blood-vessel-segmentation/discussion/475126", "author_name": "", "post_date": "2024-02-07T07:40:32.498510700Z", "votes": 6, "comment_count": 4, "views": 0, "content": "

based on the results
\n

\n

it shows that my model detects nothing (no hit and no fp) from the private dataset.

\n

this means that the textureness of private is different from public.
\n(one can confirm this via probing in post submission)

\n

then i recall that in the paper, they show images of different um per voxel (for different scan resolution).

\n

at high resolution (like voi and private dataset), there are more details. i.e. more contrast and more texture. so i think private set is a \"little shift\" to kidney1 voi dataset.

\n
\n

my mistake: foucs on \" 63.08 and 50 15.77um\" (image resolution after binning) , but actally we should focus on \"25.14 and 15.77um\" (scanning).

\n

scanning resolution for voi is not mentioned in the data page, but it is mentioned in the paper.

\n

Hence all my scaling experiments are flawed

\n

", "messages": [ { "id": "2640967", "postDate": "02/07/2024 07:40:32", "content": "

based on the results
\n

\n

it shows that my model detects nothing (no hit and no fp) from the private dataset.

\n

this means that the textureness of private is different from public.
\n(one can confirm this via probing in post submission)

\n

then i recall that in the paper, they show images of different um per voxel (for different scan resolution).

\n

at high resolution (like voi and private dataset), there are more details. i.e. more contrast and more texture. so i think private set is a \"little shift\" to kidney1 voi dataset.

\n
\n

my mistake: foucs on \" 63.08 and 50 15.77um\" (image resolution after binning) , but actally we should focus on \"25.14 and 15.77um\" (scanning).

\n

scanning resolution for voi is not mentioned in the data page, but it is mentioned in the paper.

\n

Hence all my scaling experiments are flawed

\n

", "rawMarkdown": "based on the results\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1fef6d63de4651ec1b423f85bf0d1984%2FSelection_999(4941).png?generation=1707319220577706&alt=media)\n\nit shows that my model detects nothing (no hit and no fp) from the private dataset.\n\nthis means that the textureness of private is different from public.\n(one can confirm this via probing in post submission)\n\nthen i recall that in the paper, they show images of different um per voxel (for different scan resolution).\n\nat high resolution (like voi and private dataset), there are more details. i.e. more contrast and more texture. so i think private set is a \"little shift\" to kidney1 voi dataset.\n\n---\n\n**my mistake: foucs on \" 63.08 and 50 15.77um\" (image resolution after binning) , but actally we should focus on \"25.14 and 15.77um\" (scanning).**\n\nscanning resolution for voi is not mentioned in the data page, but it is mentioned in the paper.\n\nHence all my scaling experiments are flawed\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fea7ba4bf2632ea073800aeb308aef726%2FSelection_999(4932).png?generation=1707293680693145&alt=media)", "votes": null }, { "id": "2640992", "postDate": "02/07/2024 08:06:24", "content": "

possible evidence

\n
    \n
  1. @forcewithme
    \n[3rd Place solution] Refine from Sparse to Dense
    2. \n
\n

\n

p=1.0 !!!!!!

", "rawMarkdown": "possible evidence\n1. @forcewithme \n[3rd Place solution] Refine from Sparse to Dense\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb18beebea322ff8e7dec58baa094dd9d%2FSelection_999(4934).png?generation=1707293152178917&alt=media)\n\np=1.0 !!!!!!", "votes": null }, { "id": "2640994", "postDate": "02/07/2024 08:08:58", "content": "

For me, the intensity augmentation worked well for the 2d model and 2.5d model, while for the 3d model, the degradation was both on CV and LB.

", "rawMarkdown": "For me, the intensity augmentation worked well for the 2d model and 2.5d model, while for the 3d model, the degradation was both on CV and LB.", "votes": null }, { "id": "2640998", "postDate": "02/07/2024 08:12:07", "content": "

thanks!

\n

it work for me. i note that this is the most important augmentation
\nbut i was held back and use p=0.8 and then back to p=0.5

\n

@igorkrashenyi
\n4th place solution. Boundary DoU Loss is all you need!

\n

", "rawMarkdown": "thanks!\n\nit work for me. i note that this is the most important augmentation\nbut i was held back and use p=0.8 and then back to p=0.5 \n\n@igorkrashenyi \n4th place solution. Boundary DoU Loss is all you need!\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6916ac319d47a61479fe1ab6557cc3c7%2FSelection_999(4936).png?generation=1707293410630521&alt=media)", "votes": null }, { "id": "2641002", "postDate": "02/07/2024 08:16:16", "content": "

@hengck23 Yeah, it's for the 2d and 2.5d cases.
\nI've tried to do the same for the 3d case but faced performance degradation.

", "rawMarkdown": "hengck23 Yeah, it's for the 2d and 2.5d cases. \nI've tried to do the same for the 3d case but faced performance degradation.", "votes": null } ], "comments": [ { "id": 2640992, "author_name": "hengck23", "author_url": "", "post_date": "02/07/2024 08:06:24", "content": "

possible evidence

\n
    \n
  1. @forcewithme
    \n[3rd Place solution] Refine from Sparse to Dense
    2. \n
\n

\n

p=1.0 !!!!!!

", "votes": null, "replies": [ { "id": 2640994, "author_name": "igorkrashenyi", "author_url": "", "post_date": "02/07/2024 08:08:58", "content": "

For me, the intensity augmentation worked well for the 2d model and 2.5d model, while for the 3d model, the degradation was both on CV and LB.

", "votes": null, "replies": [ { "id": 2640998, "author_name": "hengck23", "author_url": "", "post_date": "02/07/2024 08:12:07", "content": "

thanks!

\n

it work for me. i note that this is the most important augmentation
\nbut i was held back and use p=0.8 and then back to p=0.5

\n

@igorkrashenyi
\n4th place solution. Boundary DoU Loss is all you need!

\n

", "votes": null, "replies": [ { "id": 2641002, "author_name": "igorkrashenyi", "author_url": "", "post_date": "02/07/2024 08:16:16", "content": "

@hengck23 Yeah, it's for the 2d and 2.5d cases.
\nI've tried to do the same for the 3d case but faced performance degradation.

", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2640967": "based on the results\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1fef6d63de4651ec1b423f85bf0d1984%2FSelection_999(4941).png?generation=1707319220577706&alt=media)\n\nit shows that my model detects nothing (no hit and no fp) from the private dataset.\n\nthis means that the textureness of private is different from public.\n(one can confirm this via probing in post submission)\n\nthen i recall that in the paper, they show images of different um per voxel (for different scan resolution).\n\nat high resolution (like voi and private dataset), there are more details. i.e. more contrast and more texture. so i think private set is a \"little shift\" to kidney1 voi dataset.\n\n---\n\n**my mistake: foucs on \" 63.08 and 50 15.77um\" (image resolution after binning) , but actally we should focus on \"25.14 and 15.77um\" (scanning).**\n\nscanning resolution for voi is not mentioned in the data page, but it is mentioned in the paper.\n\nHence all my scaling experiments are flawed\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fea7ba4bf2632ea073800aeb308aef726%2FSelection_999(4932).png?generation=1707293680693145&alt=media)", "2640992": "possible evidence\n1. @forcewithme \n[3rd Place solution] Refine from Sparse to Dense\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb18beebea322ff8e7dec58baa094dd9d%2FSelection_999(4934).png?generation=1707293152178917&alt=media)\n\np=1.0 !!!!!!", "2640994": "For me, the intensity augmentation worked well for the 2d model and 2.5d model, while for the 3d model, the degradation was both on CV and LB.", "2640998": "thanks!\n\nit work for me. i note that this is the most important augmentation\nbut i was held back and use p=0.8 and then back to p=0.5 \n\n@igorkrashenyi \n4th place solution. Boundary DoU Loss is all you need!\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6916ac319d47a61479fe1ab6557cc3c7%2FSelection_999(4936).png?generation=1707293410630521&alt=media)", "2641002": "hengck23 Yeah, it's for the 2d and 2.5d cases. \nI've tried to do the same for the 3d case but faced performance degradation." }, "source": "meta" }