{ "id": 349378, "title": "Patch images vs whole images", "url": "/competitions/hubmap-organ-segmentation/discussion/349378", "author_name": "DDCV Lab", "post_date": "2022-09-01T07:36:22.429000", "votes": 2, "comment_count": 5, "views": 0, "content": "

I have tried trainig with whole images (resize to 768*768) and got CV: 0.83 LB: 0.65, still a large gap between CV and LB
\nDoes tiled images narrow the gap between CV and LB significantly? I have also tried stain normalization, but it seems doesn't improve a lot.

", "messages": [ { "id": 1931587, "postDate": "2022-09-08T20:03:58.207Z", "content": "

In my results, tiled dataset is better.
\n\"\"

", "rawMarkdown": "In my results, tiled dataset is better.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10229483%2F56d64e5f20ad378efb784e8031ee9eef%2F2022-09-09%20050141.png?generation=1662667434515406&alt=media)", "votes": 1 }, { "id": 1921986, "postDate": "2022-09-01T07:36:22.430Z", "content": "

I have tried trainig with whole images (resize to 768*768) and got CV: 0.83 LB: 0.65, still a large gap between CV and LB
\nDoes tiled images narrow the gap between CV and LB significantly? I have also tried stain normalization, but it seems doesn't improve a lot.

", "rawMarkdown": "I have tried trainig with whole images (resize to 768*768) and got CV: 0.83 LB: 0.65, still a large gap between CV and LB\nDoes tiled images narrow the gap between CV and LB significantly? I have also tried stain normalization, but it seems doesn't improve a lot.\n", "votes": 2 }, { "id": 1926701, "postDate": "2022-09-05T03:38:21.187Z", "content": "

I'm not getting any results on any technique.
\npatches or resized images
\ntraining from scratcch, pre-trained models, effnet backbones…
\nwe aren't even able to get visually good results
\nlet alone think of calculating the CV

", "rawMarkdown": "I'm not getting any results on any technique.\npatches or resized images\ntraining from scratcch, pre-trained models, effnet backbones...\nwe aren't even able to get visually good results\nlet alone think of calculating the CV" }, { "id": 1924296, "postDate": "2022-09-02T23:13:50.023Z", "content": "

I'm also having large split problems. You used KFold right? What models did you use?

\n

I used single fold UNet efficientB7 backbone, got 77.78% on 10% val data but 0.48 on LB…. But it is a baseline anyways.

", "rawMarkdown": "I'm also having large split problems. You used KFold right? What models did you use?\n\nI used single fold UNet efficientB7 backbone, got 77.78% on 10% val data but 0.48 on LB.... But it is a baseline anyways.", "replies": [ { "id": 1924416, "postDate": "2022-09-03T03:18:53.940Z", "content": "

I'm now using deeplabV3 + b4 and get CV 0.83 LB 0.71, which is still a large gap between CV and LB

", "rawMarkdown": "I'm now using deeplabV3 + b4 and get CV 0.83 LB 0.71, which is still a large gap between CV and LB", "votes": 1 } ] }, { "id": 1931585, "postDate": "2022-09-08T20:02:30.053Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1931587, "author_name": "Jiseong Ok", "author_url": "", "post_date": "2022-09-08T20:03:58.207000", "content": "

In my results, tiled dataset is better.
\n\"\"

", "votes": 1, "replies": [] }, { "id": 1926701, "author_name": "Bhavesh Jain", "author_url": "", "post_date": "2022-09-05T03:38:21.187000", "content": "

I'm not getting any results on any technique.
\npatches or resized images
\ntraining from scratcch, pre-trained models, effnet backbones…
\nwe aren't even able to get visually good results
\nlet alone think of calculating the CV

", "votes": 0, "replies": [] }, { "id": 1924296, "author_name": "Daniel Wang", "author_url": "", "post_date": "2022-09-02T23:13:50.023000", "content": "

I'm also having large split problems. You used KFold right? What models did you use?

\n

I used single fold UNet efficientB7 backbone, got 77.78% on 10% val data but 0.48 on LB…. But it is a baseline anyways.

", "votes": 0, "replies": [ { "id": 1924416, "author_name": "DDCV Lab", "author_url": "", "post_date": "2022-09-03T03:18:53.940000", "content": "

I'm now using deeplabV3 + b4 and get CV 0.83 LB 0.71, which is still a large gap between CV and LB

", "votes": 1, "replies": [] } ] }, { "id": 1931585, "author_name": "", "author_url": "", "post_date": "2022-09-08T20:02:30.053000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1931587": "In my results, tiled dataset is better.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10229483%2F56d64e5f20ad378efb784e8031ee9eef%2F2022-09-09%20050141.png?generation=1662667434515406&alt=media)", "1921986": "I have tried trainig with whole images (resize to 768*768) and got CV: 0.83 LB: 0.65, still a large gap between CV and LB\nDoes tiled images narrow the gap between CV and LB significantly? I have also tried stain normalization, but it seems doesn't improve a lot.\n", "1926701": "I'm not getting any results on any technique.\npatches or resized images\ntraining from scratcch, pre-trained models, effnet backbones...\nwe aren't even able to get visually good results\nlet alone think of calculating the CV", "1924296": "I'm also having large split problems. You used KFold right? What models did you use?\n\nI used single fold UNet efficientB7 backbone, got 77.78% on 10% val data but 0.48 on LB.... But it is a baseline anyways.", "1931585": "" } }