{ "id": 421579, "title": "What are we missing out on?", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/421579", "author_name": "", "post_date": "2023-07-05T22:39:59.687179200Z", "votes": 5, "comment_count": 4, "views": 0, "content": "

There has been a lot of different models, approaches out there for this competition in the past month and I too followed along them to improve. But seeing the leaderboard I assume there ought to be room for improvement… What can it be? Are the predictions not enough or inaccurate or not precise enough. Can anyone guide a bit?

", "messages": [ { "id": "2331985", "postDate": "07/05/2023 22:39:59", "content": "

There has been a lot of different models, approaches out there for this competition in the past month and I too followed along them to improve. But seeing the leaderboard I assume there ought to be room for improvement… What can it be? Are the predictions not enough or inaccurate or not precise enough. Can anyone guide a bit?

", "rawMarkdown": "There has been a lot of different models, approaches out there for this competition in the past month and I too followed along them to improve. But seeing the leaderboard I assume there ought to be room for improvement... What can it be? Are the predictions not enough or inaccurate or not precise enough. Can anyone guide a bit?", "votes": null }, { "id": "2332363", "postDate": "07/06/2023 06:42:45", "content": "

preprocessing the predicited mask using mask dilation increase the LB by about 0.1

", "rawMarkdown": "preprocessing the predicited mask using mask dilation increase the LB by about 0.1", "votes": null }, { "id": "2332409", "postDate": "07/06/2023 07:21:16", "content": "

Actually I've already used that. But thanks for the recommendation

", "rawMarkdown": "Actually I've already used that. But thanks for the recommendation", "votes": null }, { "id": "2332885", "postDate": "07/06/2023 13:37:14", "content": "

Unlike ICR, there is less discussion, and I feel that it is becoming a challenge for individuals. Also, it is true that the top LB score is almost stagnant, and it may not change dramatically in the future.

\n

I've seen most of the discussions so far, and I'll summarize what I'm currently considering.

\n

・Dilation is also effective for test data
\n・Segmentation models including blending are under consideration (YOLOv7, YOLOv8, MaskR-CNN, etc.)
\n・How to divide the training data set
\n・Change inference size
\n・Private test set trends
\n・Choose 2 submissions
\n・Other tricks

", "rawMarkdown": "Unlike ICR, there is less discussion, and I feel that it is becoming a challenge for individuals. Also, it is true that the top LB score is almost stagnant, and it may not change dramatically in the future.\n\nI've seen most of the discussions so far, and I'll summarize what I'm currently considering.\n\n・Dilation is also effective for test data\n・Segmentation models including blending are under consideration (YOLOv7, YOLOv8, MaskR-CNN, etc.)\n・How to divide the training data set\n・Change inference size\n・Private test set trends\n・Choose 2 submissions\n・Other tricks", "votes": null }, { "id": "2332967", "postDate": "07/06/2023 15:05:12", "content": "

I guess changing the inference size is the only thing left to try

", "rawMarkdown": "I guess changing the inference size is the only thing left to try", "votes": null } ], "comments": [ { "id": 2332363, "author_name": "ahmedabdelhamied2501", "author_url": "", "post_date": "07/06/2023 06:42:45", "content": "

preprocessing the predicited mask using mask dilation increase the LB by about 0.1

", "votes": null, "replies": [ { "id": 2332409, "author_name": "bhavesjain", "author_url": "", "post_date": "07/06/2023 07:21:16", "content": "

Actually I've already used that. But thanks for the recommendation

", "votes": null, "replies": [] } ] }, { "id": 2332885, "author_name": "mitsuyasuhoshino", "author_url": "", "post_date": "07/06/2023 13:37:14", "content": "

Unlike ICR, there is less discussion, and I feel that it is becoming a challenge for individuals. Also, it is true that the top LB score is almost stagnant, and it may not change dramatically in the future.

\n

I've seen most of the discussions so far, and I'll summarize what I'm currently considering.

\n

・Dilation is also effective for test data
\n・Segmentation models including blending are under consideration (YOLOv7, YOLOv8, MaskR-CNN, etc.)
\n・How to divide the training data set
\n・Change inference size
\n・Private test set trends
\n・Choose 2 submissions
\n・Other tricks

", "votes": null, "replies": [ { "id": 2332967, "author_name": "bhavesjain", "author_url": "", "post_date": "07/06/2023 15:05:12", "content": "

I guess changing the inference size is the only thing left to try

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2331985": "There has been a lot of different models, approaches out there for this competition in the past month and I too followed along them to improve. But seeing the leaderboard I assume there ought to be room for improvement... What can it be? Are the predictions not enough or inaccurate or not precise enough. Can anyone guide a bit?", "2332363": "preprocessing the predicited mask using mask dilation increase the LB by about 0.1", "2332409": "Actually I've already used that. But thanks for the recommendation", "2332885": "Unlike ICR, there is less discussion, and I feel that it is becoming a challenge for individuals. Also, it is true that the top LB score is almost stagnant, and it may not change dramatically in the future.\n\nI've seen most of the discussions so far, and I'll summarize what I'm currently considering.\n\n・Dilation is also effective for test data\n・Segmentation models including blending are under consideration (YOLOv7, YOLOv8, MaskR-CNN, etc.)\n・How to divide the training data set\n・Change inference size\n・Private test set trends\n・Choose 2 submissions\n・Other tricks", "2332967": "I guess changing the inference size is the only thing left to try" }, "source": "meta" }