How do I learn real graphs (skeletons) from training on approximate (KNNGraph) graphs?

[davodogster]

Hi guys,

This is my PyG dataset format:

[rusty1s]

Can you share some insights on what you want to do concretely and how X and Y relate to each other (since their number of nodes vary)?

[davodogster]

Hi @rusty1s

My data looks likes this.
I have 1. PCD (input) 2. Skeleton(output). 3. Mesh (green) For 10000+ trees.

(Note: these three files are not the same sample sorry)

I'd like to train a GNN to learn skeletons from point clouds. As above - skeletons have different number of vertices/pos than the number of points in the point cloud. Do you know if it's possible to train such a network in PyG. Does X_pos.shape[0] have to equal Y_pos.shape[0] and X_edge_index.shape[1] have to equal Y_edge_index.shape[1]. Its a very useful real world application if we can get this to work. Thanks, Sam

[rusty1s]

There does not exist such kind of example in PyG yet, I am sorry. Do you know of any papers that tackle this kind of problem?

[davodogster]

Hi @rusty1s thanks for your response. This one appears closest: https://onlinelibrary.wiley.com/doi/epdf/10.1111/cgf.14151
This is also similar: https://github.com/clinplayer/Point2Skeleton

This is related but for mesh only (ignore the skeleton) https://github.com/weixk2015/DHSP3D

Do you think it's possible to implement in PyG?

[davodogster]

Hi @rusty1s do you think the PCD to mesh or PCD to skeleton is possible with GNN deep learning? e.g https://github.com/clinplayer/Point2Skeleton

Thanks

[rusty1s]

Sorry for the late reply here. Point2Skeleton is a cool paper :) Wouldn't this be something that is directly applicable in your case? The paper uses a GNN (i.e. PointNet++) to learn the skeletal points already.