![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8486392%2Fef5d758097d9c2e9a3650f5f9b5ca44b%2FScreenshot%202025-03-22%20192907.png?generation=1742651995024863&alt=media\")
in code, many codes come from yolo detection for 2d image. Can this be solved by 3d detecting/classifying models?
", "messages": [ { "id": "3152702", "postDate": "03/18/2025 04:46:49", "content": "in code, many codes come from yolo detection for 2d image. Can this be solved by 3d detecting/classifying models?
", "rawMarkdown": "in code, many codes come from yolo detection for 2d image. Can this be solved by 3d detecting/classifying models?", "votes": null }, { "id": "3152714", "postDate": "03/18/2025 05:10:24", "content": "For example, with 3D U-Net, I've conducted some trials and found that the inference time exceeds 12 hours. If this issue can be resolved, then 3D U-Net can be used. Currently, I'm still in the process of trying to figure it out.
", "rawMarkdown": "For example, with 3D U-Net, I've conducted some trials and found that the inference time exceeds 12 hours. If this issue can be resolved, then 3D U-Net can be used. Currently, I'm still in the process of trying to figure it out.", "votes": null }, { "id": "3152716", "postDate": "03/18/2025 05:11:12", "content": "I'm going to try a two-stage model.
", "rawMarkdown": "I'm going to try a two-stage model.", "votes": null }, { "id": "3152725", "postDate": "03/18/2025 05:24:23", "content": "I agree only using one stage will consume too many memories since pixels are too many in 3d compared to 2d. two stage might be worth to try. I will also consider that approach thanks
", "rawMarkdown": "I agree only using one stage will consume too many memories since pixels are too many in 3d compared to 2d. two stage might be worth to try. I will also consider that approach thanks", "votes": null }, { "id": "3152772", "postDate": "03/18/2025 06:29:56", "content": "Maybe we can try 2-stage model. One of model is applied to classifying the slice number. I used to participate in RSNA2024, and most of people in this competition use this method.
", "rawMarkdown": "Maybe we can try 2-stage model. One of model is applied to classifying the slice number. I used to participate in RSNA2024, and most of people in this competition use this method.", "votes": null }, { "id": "3152795", "postDate": "03/18/2025 07:03:09", "content": "In the perspective of decreasing complexity, classifying the slice number is a good way. I want to further figure out whether using 3d reconstruction for images in local area(images remaining after first stage) is useful for optimizing the model pipeline
", "rawMarkdown": "In the perspective of decreasing complexity, classifying the slice number is a good way. I want to further figure out whether using 3d reconstruction for images in local area(images remaining after first stage) is useful for optimizing the model pipeline", "votes": null }, { "id": "3152865", "postDate": "03/18/2025 08:15:18", "content": "I cannot make 3d models work in my experiments.
", "rawMarkdown": "I cannot make 3d models work in my experiments.", "votes": null }, { "id": "3152870", "postDate": "03/18/2025 08:18:56", "content": "since we don't know the exact camera's intrinsic/extrinsic parameters?
", "rawMarkdown": "since we don't know the exact camera's intrinsic/extrinsic parameters?", "votes": null }, { "id": "3152874", "postDate": "03/18/2025 08:20:11", "content": "My 3D Unet model doesn't work at all😬 , I tried different loss, model backend, patch size and target sphere. I'm probably doing something wrong
", "rawMarkdown": "My 3D Unet model doesn't work at all😬 , I tried different loss, model backend, patch size and target sphere. I'm probably doing something wrong", "votes": null }, { "id": "3152883", "postDate": "03/18/2025 08:27:51", "content": "If the problem is wrong 3d input data generating(wrong reconstruction), then what about eliminating some slices instead of using all the slices. I'm new to this so I don't know well.
", "rawMarkdown": "If the problem is wrong 3d input data generating(wrong reconstruction), then what about eliminating some slices instead of using all the slices. I'm new to this so I don't know well.", "votes": null }, { "id": "3152888", "postDate": "03/18/2025 08:33:11", "content": "As far as I know, medical dataset are not like the way reality is. So, depth estimation is really hard and because of that, it is hard to generate accurate 3d compared to objects existing in reality
", "rawMarkdown": "As far as I know, medical dataset are not like the way reality is. So, depth estimation is really hard and because of that, it is hard to generate accurate 3d compared to objects existing in reality", "votes": null }, { "id": "3156764", "postDate": "03/22/2025 14:07:38", "content": "yes but not work !!!!!!!!!!
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8486392%2Fccf2d1c4238480048234f330951490fc%2FScreenshot%202025-03-22%20193707.png?generation=1742652454675790&alt=media\")
@i2nfinit3y that's great insight thanks
", "rawMarkdown": "i2nfinit3y that's great insight thanks", "votes": null } ], "comments": [ { "id": 3152714, "author_name": "switch9527", "author_url": "", "post_date": "03/18/2025 05:10:24", "content": "For example, with 3D U-Net, I've conducted some trials and found that the inference time exceeds 12 hours. If this issue can be resolved, then 3D U-Net can be used. Currently, I'm still in the process of trying to figure it out.
", "votes": null, "replies": [ { "id": 3152716, "author_name": "switch9527", "author_url": "", "post_date": "03/18/2025 05:11:12", "content": "I'm going to try a two-stage model.
", "votes": null, "replies": [ { "id": 3156764, "author_name": "prothomeshmistry", "author_url": "", "post_date": "03/22/2025 14:07:38", "content": "yes but not work !!!!!!!!!!
I agree only using one stage will consume too many memories since pixels are too many in 3d compared to 2d. two stage might be worth to try. I will also consider that approach thanks
", "votes": null, "replies": [] } ] }, { "id": 3152772, "author_name": "i2nfinit3y", "author_url": "", "post_date": "03/18/2025 06:29:56", "content": "Maybe we can try 2-stage model. One of model is applied to classifying the slice number. I used to participate in RSNA2024, and most of people in this competition use this method.
", "votes": null, "replies": [ { "id": 3152795, "author_name": "boochanggyu", "author_url": "", "post_date": "03/18/2025 07:03:09", "content": "In the perspective of decreasing complexity, classifying the slice number is a good way. I want to further figure out whether using 3d reconstruction for images in local area(images remaining after first stage) is useful for optimizing the model pipeline
", "votes": null, "replies": [] }, { "id": 3158617, "author_name": "tom99763", "author_url": "", "post_date": "03/24/2025 17:51:25", "content": "@i2nfinit3y that's great insight thanks
", "votes": null, "replies": [] } ] }, { "id": 3152865, "author_name": "zacchaeus", "author_url": "", "post_date": "03/18/2025 08:15:18", "content": "I cannot make 3d models work in my experiments.
", "votes": null, "replies": [ { "id": 3152870, "author_name": "boochanggyu", "author_url": "", "post_date": "03/18/2025 08:18:56", "content": "since we don't know the exact camera's intrinsic/extrinsic parameters?
", "votes": null, "replies": [] } ] }, { "id": 3152874, "author_name": "evgeny000", "author_url": "", "post_date": "03/18/2025 08:20:11", "content": "My 3D Unet model doesn't work at all😬 , I tried different loss, model backend, patch size and target sphere. I'm probably doing something wrong
", "votes": null, "replies": [ { "id": 3152883, "author_name": "boochanggyu", "author_url": "", "post_date": "03/18/2025 08:27:51", "content": "If the problem is wrong 3d input data generating(wrong reconstruction), then what about eliminating some slices instead of using all the slices. I'm new to this so I don't know well.
", "votes": null, "replies": [] }, { "id": 3152888, "author_name": "boochanggyu", "author_url": "", "post_date": "03/18/2025 08:33:11", "content": "As far as I know, medical dataset are not like the way reality is. So, depth estimation is really hard and because of that, it is hard to generate accurate 3d compared to objects existing in reality
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3152702": "in code, many codes come from yolo detection for 2d image. Can this be solved by 3d detecting/classifying models?", "3152714": "For example, with 3D U-Net, I've conducted some trials and found that the inference time exceeds 12 hours. If this issue can be resolved, then 3D U-Net can be used. Currently, I'm still in the process of trying to figure it out.", "3152716": "I'm going to try a two-stage model.", "3152725": "I agree only using one stage will consume too many memories since pixels are too many in 3d compared to 2d. two stage might be worth to try. I will also consider that approach thanks", "3152772": "Maybe we can try 2-stage model. One of model is applied to classifying the slice number. I used to participate in RSNA2024, and most of people in this competition use this method.", "3152795": "In the perspective of decreasing complexity, classifying the slice number is a good way. I want to further figure out whether using 3d reconstruction for images in local area(images remaining after first stage) is useful for optimizing the model pipeline", "3152865": "I cannot make 3d models work in my experiments.", "3152870": "since we don't know the exact camera's intrinsic/extrinsic parameters?", "3152874": "My 3D Unet model doesn't work at all😬 , I tried different loss, model backend, patch size and target sphere. I'm probably doing something wrong", "3152883": "If the problem is wrong 3d input data generating(wrong reconstruction), then what about eliminating some slices instead of using all the slices. I'm new to this so I don't know well.", "3152888": "As far as I know, medical dataset are not like the way reality is. So, depth estimation is really hard and because of that, it is hard to generate accurate 3d compared to objects existing in reality", "3156764": "yes but not work !!!!!!!!!! \n\n", "3158617": "i2nfinit3y that's great insight thanks" }, "source": "meta" }