![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff396f30e1f184b7b7bc79fe179abfed5%2FSelection_1869.png?generation=1766672352357356&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fcd28771b0afe9015e0822dbfad5b473b%2FSelection_1868.png?generation=1766672364492222&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff26eacaa5680a47fd5df183d516cfd5f%2FSelection_1867.png?generation=1766672404030337&alt=media\")
other comments:
\nBut since the thickness is now near constant, it should become easier in post processing to fix minor gaps.
", "rawMarkdown": "But since the thickness is now near constant, it should become easier in post processing to fix minor gaps.", "votes": null }, { "id": "3381809", "postDate": "12/25/2025 15:21:46", "content": "For the most part when deciding whether or not to remove the single isolated sheets we went with the decision to remove them when fixing other topological errors. Their presence is a bit challenging from an evaluation standpoint as well considering they are on both sides surrounded by ignore label increasing the chances of spurious components from nearby sheets invading.
\nOn top of both of those , from a pure visual inspection standpoint it’s very difficult to determine if you are “missing sheets” on either side of a single sheet in a prediction , and becomes much easier the more you have
\nIn my opinion at least it was “cleaner” to keep primarily areas which have more than one (ideally 3) sheets. The isolated ones also provide such sparse supervision that it’s hard to say if they’re doing much other than generating noise / wasting compute anyways.
", "rawMarkdown": "For the most part when deciding whether or not to remove the single isolated sheets we went with the decision to remove them when fixing other topological errors. Their presence is a bit challenging from an evaluation standpoint as well considering they are on both sides surrounded by ignore label increasing the chances of spurious components from nearby sheets invading.\n\nOn top of both of those , from a pure visual inspection standpoint it’s very difficult to determine if you are “missing sheets” on either side of a single sheet in a prediction , and becomes much easier the more you have \n\nIn my opinion at least it was “cleaner” to keep primarily areas which have more than one (ideally 3) sheets. The isolated ones also provide such sparse supervision that it’s hard to say if they’re doing much other than generating noise / wasting compute anyways.", "votes": null }, { "id": "3381878", "postDate": "12/25/2025 19:42:19", "content": "it is easier to do in 2 steps:
\n1) stage1: fix gap/topology etc in skeleton
\n2) stage2: input = sketelon+image and backprop lb metric related loss
Hmmm, i was also thinking this only. Naive segmentation --> bridge/merges fix--> now we have seperate sheets but with holes --> holes fix --> perfectly seperated sheets to work further with
", "rawMarkdown": "Hmmm, i was also thinking this only. Naive segmentation --> bridge/merges fix--> now we have seperate sheets but with holes --> holes fix --> perfectly seperated sheets to work further with", "votes": null }, { "id": "3382332", "postDate": "12/27/2025 08:14:39", "content": "voxel, I think? @hengck23
", "rawMarkdown": "voxel, I think? @hengck23", "votes": null }, { "id": "3382344", "postDate": "12/27/2025 09:07:41", "content": "update:
\nother comments:
\nBut since the thickness is now near constant, it should become easier in post processing to fix minor gaps.
", "votes": null, "replies": [ { "id": 3381878, "author_name": "hengck23", "author_url": "", "post_date": "12/25/2025 19:42:19", "content": "it is easier to do in 2 steps:
\n1) stage1: fix gap/topology etc in skeleton
\n2) stage2: input = sketelon+image and backprop lb metric related loss
Hmmm, i was also thinking this only. Naive segmentation --> bridge/merges fix--> now we have seperate sheets but with holes --> holes fix --> perfectly seperated sheets to work further with
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3381809, "author_name": "seanjohnsonsp", "author_url": "", "post_date": "12/25/2025 15:21:46", "content": "For the most part when deciding whether or not to remove the single isolated sheets we went with the decision to remove them when fixing other topological errors. Their presence is a bit challenging from an evaluation standpoint as well considering they are on both sides surrounded by ignore label increasing the chances of spurious components from nearby sheets invading.
\nOn top of both of those , from a pure visual inspection standpoint it’s very difficult to determine if you are “missing sheets” on either side of a single sheet in a prediction , and becomes much easier the more you have
\nIn my opinion at least it was “cleaner” to keep primarily areas which have more than one (ideally 3) sheets. The isolated ones also provide such sparse supervision that it’s hard to say if they’re doing much other than generating noise / wasting compute anyways.
", "votes": null, "replies": [] }, { "id": 3382332, "author_name": "navneetbende", "author_url": "", "post_date": "12/27/2025 08:14:39", "content": "voxel, I think? @hengck23
", "votes": null, "replies": [] }, { "id": 3382344, "author_name": "hengck23", "author_url": "", "post_date": "12/27/2025 09:07:41", "content": "update:
\n