\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Ff68357499ee1e41080f75d356b7605b4%2FScreenshot_20250308_163545_Chrome.jpg?generation=1741741349231554&alt=media\")
I currently create a result like that. Is it possible to build a gnn and predict the 3d location? I'm struggling on the output development.
\n
I currently create a result like that. Is it possible to build a gnn and predict the 3d location? I'm struggling on the output development.
\n
How did u build this cloud? If we just combine the slices, there would be a voxel grid, no? So u don't include some points into cloud?
\nI don't really know about GNNs, but as far as i know topologic methods can be quite useful for point clouds.
", "rawMarkdown": "How did u build this cloud? If we just combine the slices, there would be a voxel grid, no? So u don't include some points into cloud?\n\nI don't really know about GNNs, but as far as i know topologic methods can be quite useful for point clouds.", "votes": null }, { "id": "3147345", "postDate": "03/12/2025 00:33:07", "content": "https://youtu.be/3apDWJWe_jg?si=eH0Oyeh0ieujep93
", "rawMarkdown": "https://youtu.be/3apDWJWe_jg?si=eH0Oyeh0ieujep93", "votes": null }, { "id": "3147352", "postDate": "03/12/2025 00:37:54", "content": "I use the open3d package to build this cloud. But the main challenge is making the ground truth.
@tom99763 thx for the link
\nhonestly, i don't really get what's the problem. The problem is in generating targets for the single point cloud? Or is it a problem to build the best point cloud (selecting brightness threshold)? Or something else?
\nWhat do you mean by 'ground truth' here?
@eikyou Build the point cloud on the 3d image->Graph construction based on knn -> feed into GNN and it outputs 0 or 1 for each node which means whether this node contains flagellar motors. My question is how to create ground truth labels?
", "rawMarkdown": "eikyou Build the point cloud on the 3d image->Graph construction based on knn -> feed into GNN and it outputs 0 or 1 for each node which means whether this node contains flagellar motors. My question is how to create ground truth labels?", "votes": null }, { "id": "3147391", "postDate": "03/12/2025 02:02:32", "content": "@tom99763 i would firstly try to assign 1 (there is a motor) for each node within the range <=1000 angstrem from the given center. Maybe it's reasonable to narrow down the range to 500(?) or 800(?). Have u tried this?
", "rawMarkdown": "tom99763 i would firstly try to assign 1 (there is a motor) for each node within the range <=1000 angstrem from the given center. Maybe it's reasonable to narrow down the range to 500(?) or 800(?). Have u tried this?", "votes": null }, { "id": "3186374", "postDate": "04/24/2025 15:12:28", "content": "The approach from PYG seems works: https://pytorch-geometric.readthedocs.io/en/latest/tutorial/point_cloud.html
", "rawMarkdown": "The approach from PYG seems works: https://pytorch-geometric.readthedocs.io/en/latest/tutorial/point_cloud.html", "votes": null } ], "comments": [ { "id": 3147281, "author_name": "eikyou", "author_url": "", "post_date": "03/11/2025 21:28:07", "content": "How did u build this cloud? If we just combine the slices, there would be a voxel grid, no? So u don't include some points into cloud?
\nI don't really know about GNNs, but as far as i know topologic methods can be quite useful for point clouds.
", "votes": null, "replies": [ { "id": 3147345, "author_name": "tom99763", "author_url": "", "post_date": "03/12/2025 00:33:07", "content": "https://youtu.be/3apDWJWe_jg?si=eH0Oyeh0ieujep93
", "votes": null, "replies": [ { "id": 3147352, "author_name": "tom99763", "author_url": "", "post_date": "03/12/2025 00:37:54", "content": "I use the open3d package to build this cloud. But the main challenge is making the ground truth.
@tom99763 thx for the link
\nhonestly, i don't really get what's the problem. The problem is in generating targets for the single point cloud? Or is it a problem to build the best point cloud (selecting brightness threshold)? Or something else?
\nWhat do you mean by 'ground truth' here?
@eikyou Build the point cloud on the 3d image->Graph construction based on knn -> feed into GNN and it outputs 0 or 1 for each node which means whether this node contains flagellar motors. My question is how to create ground truth labels?
", "votes": null, "replies": [ { "id": 3147391, "author_name": "eikyou", "author_url": "", "post_date": "03/12/2025 02:02:32", "content": "@tom99763 i would firstly try to assign 1 (there is a motor) for each node within the range <=1000 angstrem from the given center. Maybe it's reasonable to narrow down the range to 500(?) or 800(?). Have u tried this?
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3186374, "author_name": "tom99763", "author_url": "", "post_date": "04/24/2025 15:12:28", "content": "The approach from PYG seems works: https://pytorch-geometric.readthedocs.io/en/latest/tutorial/point_cloud.html
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3147058": "I currently create a result like that. Is it possible to build a gnn and predict the 3d location? I'm struggling on the output development.\n", "3147281": "How did u build this cloud? If we just combine the slices, there would be a voxel grid, no? So u don't include some points into cloud?\n\nI don't really know about GNNs, but as far as i know topologic methods can be quite useful for point clouds.", "3147345": "https://youtu.be/3apDWJWe_jg?si=eH0Oyeh0ieujep93", "3147352": "I use the `open3d` package to build this cloud. But the main challenge is making the ground truth.", "3147375": "tom99763 thx for the link\nhonestly, i don't really get what's the problem. The problem is in generating targets for the single point cloud? Or is it a problem to build the best point cloud (selecting brightness threshold)? Or something else?\nWhat do you mean by 'ground truth' here?", "3147387": "eikyou Build the point cloud on the 3d image->Graph construction based on knn -> feed into GNN and it outputs 0 or 1 for each node which means whether this node contains flagellar motors. My question is how to create ground truth labels?", "3147391": "tom99763 i would firstly try to assign 1 (there is a motor) for each node within the range <=1000 angstrem from the given center. Maybe it's reasonable to narrow down the range to 500(?) or 800(?). Have u tried this?", "3186374": "The approach from PYG seems works: https://pytorch-geometric.readthedocs.io/en/latest/tutorial/point_cloud.html" }, "source": "meta" }