search

torch-points3d / torch-points3d

Pytorch framework for doing deep learning on point clouds.
https://torch-points3d.readthedocs.io/en/latest/
Other
2.51k stars 393 forks source link

add new task : learning descriptor for point cloud registration #118

Closed humanpose1 closed 4 years ago

humanpose1 commented 4 years ago

It would be great to also have a benchmark for the descriptors. So that we could compare the different convolutions on this task (KPConv, RSConv, pointnet++....). Dataset : 3DMatch 1) we have to download the dataset (it is RGBD frames) 2) create fragments 3) find pair of similar patches from different fragments 4) (optionally some preprocessing) 5) siamese network to learn descriptors on patches

An other option is to learn descriptor on the fragment itself (like FCGF like a segmentation network) and not on patches.

nicolas-chaulet commented 4 years ago

That's a great idea! I think the starting point is to add 3d match as a torch geometric dataset and implement the relevant transformations to make it training ready (a bit like how shapenet gets processed from raw individuals files and labels into batches ready to be fed to a network). Feel free to get started on that!

nicolas-chaulet commented 4 years ago

Closing as it is done!