{ "id": 417071, "title": "Unannotated Signal in Fragment ID 1???", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/417071", "author_name": "", "post_date": "2023-06-14T04:53:03.044851300Z", "votes": null, "comment_count": 3, "views": 0, "content": "

I am afraid to be too late to make this topic.
\nI notice that my models consistently make a mistake in the upperright part of Fragment ID 1.
\n\"
\nInterestingly, I saw the models by @hengck23 and @tanakar also made the similar mistake.
\n\"
\n\"
\nI also found some signs of ink in photo and IR image of this fragment.
\n\"
\nDo anybody see similar errors in your model? Do you think it is unannotated signal?
\n@stephenrparsons Do you have any comments on it?

", "messages": [ { "id": "2301663", "postDate": "06/14/2023 04:53:03", "content": "

I am afraid to be too late to make this topic.
\nI notice that my models consistently make a mistake in the upperright part of Fragment ID 1.
\n\"
\nInterestingly, I saw the models by @hengck23 and @tanakar also made the similar mistake.
\n\"
\n\"
\nI also found some signs of ink in photo and IR image of this fragment.
\n\"
\nDo anybody see similar errors in your model? Do you think it is unannotated signal?
\n@stephenrparsons Do you have any comments on it?

", "rawMarkdown": "I am afraid to be too late to make this topic.\nI notice that my models consistently make a *mistake* in the upperright part of Fragment ID 1.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Fdcc60cfac8bbd4de9405d880660a49f7%2Fres.jpg?generation=1686717756322197&alt=media)\nInterestingly, I saw the models by @hengck23 and @tanakar also made the similar *mistake*.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Fe5eb94aea039c11334cf87117d9978f0%2Finbox_113660_b3ae8a9d4cb24ac90713e416253ee775_Selection_999(2170).png?generation=1686717879262018&alt=media)\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2F6f34aadeed4c531132e5cda98ab6d47d%2F__results___44_1.png?generation=1686717895579262&alt=media)\nI also found some signs of ink in photo and IR image of this fragment.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Ff817f029a8fc59b015b780057503965c%2Finbox_13132488_a80987e5e9973d4f338e18fb8583abee_Screen%20Shot%202023-02-20%20at%2015.51.03.png?generation=1686717975306830&alt=media)\nDo anybody see similar errors in your model? Do you think it is unannotated signal?\n@stephenrparsons Do you have any comments on it?", "votes": null }, { "id": "2301724", "postDate": "06/14/2023 06:09:06", "content": "

You can cut out dubious part and feed to your model. That might avoid label noise.

\n

I think consistently labeling this competition’s dataset is quite difficult because we have to distinguish carbon ink on the carbonized papyrus. I believe the annotator just mark as ink if darker part appeared in the IR image. Of course there can be label noise that non-ink dark part (debris on papyrus etc.) as ink labels.

", "rawMarkdown": "You can cut out dubious part and feed to your model. That might avoid label noise.\n\nI think consistently labeling this competition’s dataset is quite difficult because we have to distinguish carbon ink on the carbonized papyrus. I believe the annotator just mark as ink if darker part appeared in the IR image. Of course there can be label noise that non-ink dark part (debris on papyrus etc.) as ink labels.", "votes": null }, { "id": "2301758", "postDate": "06/14/2023 06:36:23", "content": "

check the dataset paper.
\ni think it mention about hidden fragments

", "rawMarkdown": "check the dataset paper.\ni think it mention about hidden fragments", "votes": null }, { "id": "2302127", "postDate": "06/14/2023 11:02:52", "content": "

I have also false positives in that area.

\n

\"\"

", "rawMarkdown": "I have also false positives in that area. \n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F207760%2F2176c4684a32dcd3f2b2678b1c16e6ac%2F__results___7_1.png?generation=1686740525398255&alt=media)", "votes": null } ], "comments": [ { "id": 2301724, "author_name": "tatamikenn", "author_url": "", "post_date": "06/14/2023 06:09:06", "content": "

You can cut out dubious part and feed to your model. That might avoid label noise.

\n

I think consistently labeling this competition’s dataset is quite difficult because we have to distinguish carbon ink on the carbonized papyrus. I believe the annotator just mark as ink if darker part appeared in the IR image. Of course there can be label noise that non-ink dark part (debris on papyrus etc.) as ink labels.

", "votes": null, "replies": [] }, { "id": 2301758, "author_name": "hengck23", "author_url": "", "post_date": "06/14/2023 06:36:23", "content": "

check the dataset paper.
\ni think it mention about hidden fragments

", "votes": null, "replies": [] }, { "id": 2302127, "author_name": "igorkrashenyi", "author_url": "", "post_date": "06/14/2023 11:02:52", "content": "

I have also false positives in that area.

\n

\"\"

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2301663": "I am afraid to be too late to make this topic.\nI notice that my models consistently make a *mistake* in the upperright part of Fragment ID 1.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Fdcc60cfac8bbd4de9405d880660a49f7%2Fres.jpg?generation=1686717756322197&alt=media)\nInterestingly, I saw the models by @hengck23 and @tanakar also made the similar *mistake*.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Fe5eb94aea039c11334cf87117d9978f0%2Finbox_113660_b3ae8a9d4cb24ac90713e416253ee775_Selection_999(2170).png?generation=1686717879262018&alt=media)\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2F6f34aadeed4c531132e5cda98ab6d47d%2F__results___44_1.png?generation=1686717895579262&alt=media)\nI also found some signs of ink in photo and IR image of this fragment.\n![ ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3754725%2Ff817f029a8fc59b015b780057503965c%2Finbox_13132488_a80987e5e9973d4f338e18fb8583abee_Screen%20Shot%202023-02-20%20at%2015.51.03.png?generation=1686717975306830&alt=media)\nDo anybody see similar errors in your model? Do you think it is unannotated signal?\n@stephenrparsons Do you have any comments on it?", "2301724": "You can cut out dubious part and feed to your model. That might avoid label noise.\n\nI think consistently labeling this competition’s dataset is quite difficult because we have to distinguish carbon ink on the carbonized papyrus. I believe the annotator just mark as ink if darker part appeared in the IR image. Of course there can be label noise that non-ink dark part (debris on papyrus etc.) as ink labels.", "2301758": "check the dataset paper.\ni think it mention about hidden fragments", "2302127": "I have also false positives in that area. \n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F207760%2F2176c4684a32dcd3f2b2678b1c16e6ac%2F__results___7_1.png?generation=1686740525398255&alt=media)" }, "source": "meta" }