{ "id": 414014, "title": "What magic does lb 0.77 have? why do so many people stack there?", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/414014", "author_name": "", "post_date": "2023-05-31T04:07:40.482511100Z", "votes": 10, "comment_count": 3, "views": 0, "content": "

It is interesting that some teams reached lb 0.77 a long time ago, but except for the first team, all of them were unable to break through 0.77. Is there an important thing that has not been discovered?
\n\"\"

", "messages": [ { "id": "2281661", "postDate": "05/31/2023 04:07:40", "content": "

It is interesting that some teams reached lb 0.77 a long time ago, but except for the first team, all of them were unable to break through 0.77. Is there an important thing that has not been discovered?
\n\"\"

", "rawMarkdown": "It is interesting that some teams reached lb 0.77 a long time ago, but except for the first team, all of them were unable to break through 0.77. Is there an important thing that has not been discovered?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8293010%2F739adebb420555b5531c2617e849e194%2FIMG_20230531_114323.jpg?generation=1685504905377667&alt=media)", "votes": null }, { "id": "2282628", "postDate": "05/31/2023 17:47:40", "content": "

I have feeling that difference between 0.8 and 0.77 on public leaderboard might actually be something like border clipping as described here: https://www.kaggle.com/code/elvenmonk/training-no-trimming

", "rawMarkdown": "I have feeling that difference between 0.8 and 0.77 on public leaderboard might actually be something like border clipping as described here: https://www.kaggle.com/code/elvenmonk/training-no-trimming", "votes": null }, { "id": "2287023", "postDate": "06/04/2023 04:12:40", "content": "

it is easy to prove that

\n

1) using border clipping, etc divide fragment to 2 region: A (interior) and B (removed exterior)
\n2) you have a model and subitted on region A+B with ln score s0
\n3) make submission of the same model for A only (s1) and B only (s2)
\n4) you can reverse enginner and compute how many ink pixels (truth or predicted) are in region A and B, and how prediction in different region affects score

", "rawMarkdown": "it is easy to prove that\n\n1) using border clipping, etc divide fragment to 2 region: A (interior) and B (removed exterior)\n2) you have a model and subitted on region A+B with ln score s0\n3) make submission of the same model for A only (s1) and B only (s2)\n4) you can reverse enginner and compute how many ink pixels (truth or predicted) are in region A and B, and how prediction in different region affects score", "votes": null }, { "id": "2290180", "postDate": "06/06/2023 15:55:35", "content": "

@yoyobar , I want to share an idea. Each text is wroten row by row. I think we can extract info from the right side of the page. Each row should have a different pattern with or without character.

", "rawMarkdown": "yoyobar , I want to share an idea. Each text is wroten row by row. I think we can extract info from the right side of the page. Each row should have a different pattern with or without character.", "votes": null } ], "comments": [ { "id": 2282628, "author_name": "elvenmonk", "author_url": "", "post_date": "05/31/2023 17:47:40", "content": "

I have feeling that difference between 0.8 and 0.77 on public leaderboard might actually be something like border clipping as described here: https://www.kaggle.com/code/elvenmonk/training-no-trimming

", "votes": null, "replies": [ { "id": 2287023, "author_name": "hengck23", "author_url": "", "post_date": "06/04/2023 04:12:40", "content": "

it is easy to prove that

\n

1) using border clipping, etc divide fragment to 2 region: A (interior) and B (removed exterior)
\n2) you have a model and subitted on region A+B with ln score s0
\n3) make submission of the same model for A only (s1) and B only (s2)
\n4) you can reverse enginner and compute how many ink pixels (truth or predicted) are in region A and B, and how prediction in different region affects score

", "votes": null, "replies": [] } ] }, { "id": 2290180, "author_name": "dgi1995", "author_url": "", "post_date": "06/06/2023 15:55:35", "content": "

@yoyobar , I want to share an idea. Each text is wroten row by row. I think we can extract info from the right side of the page. Each row should have a different pattern with or without character.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2281661": "It is interesting that some teams reached lb 0.77 a long time ago, but except for the first team, all of them were unable to break through 0.77. Is there an important thing that has not been discovered?\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8293010%2F739adebb420555b5531c2617e849e194%2FIMG_20230531_114323.jpg?generation=1685504905377667&alt=media)", "2282628": "I have feeling that difference between 0.8 and 0.77 on public leaderboard might actually be something like border clipping as described here: https://www.kaggle.com/code/elvenmonk/training-no-trimming", "2287023": "it is easy to prove that\n\n1) using border clipping, etc divide fragment to 2 region: A (interior) and B (removed exterior)\n2) you have a model and subitted on region A+B with ln score s0\n3) make submission of the same model for A only (s1) and B only (s2)\n4) you can reverse enginner and compute how many ink pixels (truth or predicted) are in region A and B, and how prediction in different region affects score", "2290180": "yoyobar , I want to share an idea. Each text is wroten row by row. I think we can extract info from the right side of the page. Each row should have a different pattern with or without character." }, "source": "meta" }