According to my observation, there are mainly two solutions to this problem. The first solution is going through every voxel (small block splitted) in the scan result and predict whether its center contains ink, as done in the sample submission. Another solution would be splitting the scan into larger blocks, for example 65x512x512, and perform a pixel-wise classification. Then, these blocks are assembled together to form a prediction. What are some advantages and disadvantages for these two approaches? Thank you very much for sharing ideas.
", "messages": [ { "id": "2201039", "postDate": "03/29/2023 02:16:11", "content": "According to my observation, there are mainly two solutions to this problem. The first solution is going through every voxel (small block splitted) in the scan result and predict whether its center contains ink, as done in the sample submission. Another solution would be splitting the scan into larger blocks, for example 65x512x512, and perform a pixel-wise classification. Then, these blocks are assembled together to form a prediction. What are some advantages and disadvantages for these two approaches? Thank you very much for sharing ideas.
", "rawMarkdown": "According to my observation, there are mainly two solutions to this problem. The first solution is going through every voxel (small block splitted) in the scan result and predict whether its center contains ink, as done in the sample submission. Another solution would be splitting the scan into larger blocks, for example 65x512x512, and perform a pixel-wise classification. Then, these blocks are assembled together to form a prediction. What are some advantages and disadvantages for these two approaches? Thank you very much for sharing ideas.", "votes": null }, { "id": "2206817", "postDate": "04/02/2023 22:29:27", "content": "I am checking UNet-like and it can take (fits on a GPU) about 4 crops per batch, every crop is 65x96x96 in size. It does not make sense to train one pixel at a time. It looks like a standard segmentation problem to me (just need to preprocess 3d first).
", "rawMarkdown": "I am checking UNet-like and it can take (fits on a GPU) about 4 crops per batch, every crop is 65x96x96 in size. It does not make sense to train one pixel at a time. It looks like a standard segmentation problem to me (just need to preprocess 3d first).", "votes": null }, { "id": "2207336", "postDate": "04/03/2023 10:45:11", "content": "At this point both approaches seems to be able to work. Advantages of performing classification (when you predict only one value for example if center contains ink or is there any ink in the block) are simpler training procedure and simpler architecture - you can train any binary classifier.
\nFor the segmentation approach you have a bit more complex architectures, in my case longer training times.
", "rawMarkdown": "At this point both approaches seems to be able to work. Advantages of performing classification (when you predict only one value for example if center contains ink or is there any ink in the block) are simpler training procedure and simpler architecture - you can train any binary classifier.\n\nFor the segmentation approach you have a bit more complex architectures, in my case longer training times.", "votes": null } ], "comments": [ { "id": 2206817, "author_name": "dmitrykonovalov", "author_url": "", "post_date": "04/02/2023 22:29:27", "content": "I am checking UNet-like and it can take (fits on a GPU) about 4 crops per batch, every crop is 65x96x96 in size. It does not make sense to train one pixel at a time. It looks like a standard segmentation problem to me (just need to preprocess 3d first).
", "votes": null, "replies": [] }, { "id": 2207336, "author_name": "danieliusk", "author_url": "", "post_date": "04/03/2023 10:45:11", "content": "At this point both approaches seems to be able to work. Advantages of performing classification (when you predict only one value for example if center contains ink or is there any ink in the block) are simpler training procedure and simpler architecture - you can train any binary classifier.
\nFor the segmentation approach you have a bit more complex architectures, in my case longer training times.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2201039": "According to my observation, there are mainly two solutions to this problem. The first solution is going through every voxel (small block splitted) in the scan result and predict whether its center contains ink, as done in the sample submission. Another solution would be splitting the scan into larger blocks, for example 65x512x512, and perform a pixel-wise classification. Then, these blocks are assembled together to form a prediction. What are some advantages and disadvantages for these two approaches? Thank you very much for sharing ideas.", "2206817": "I am checking UNet-like and it can take (fits on a GPU) about 4 crops per batch, every crop is 65x96x96 in size. It does not make sense to train one pixel at a time. It looks like a standard segmentation problem to me (just need to preprocess 3d first).", "2207336": "At this point both approaches seems to be able to work. Advantages of performing classification (when you predict only one value for example if center contains ink or is there any ink in the block) are simpler training procedure and simpler architecture - you can train any binary classifier.\n\nFor the segmentation approach you have a bit more complex architectures, in my case longer training times." }, "source": "meta" }