Looking for Examples of Newer Pointcloud Architectures

[fishbotics]

Hi all,

I'm wondering if anyone has implemented any of the new-ish point cloud processing architecture, such as PointNeXt or DeLA in PyG. Pointnet++ has been around for a while and there are bold claims in some recent papers of way outperforming its performance.

This discussion doesn't have to be 100% about PointNeXt, but it seems like a fairly reasonable possibility that someone out there has implemented it. The changes a few small things from Pointnet++

0. The points are all mapped via an MLP into a higher dimension at the beginning
1. The relative positions between points that are passed into the MLP are scaled according to the radius of that layer (well, multiplied by the inverse of the radius).
2. The MLPs after the SA layers are swapped out for an Inverted Residual MLP mlp (their term, not mine) does a few things, but mainly adds a skip connection in the MLP to prevent disappearing gradients.

If anyone has ideas on how to throw this together or pointers to other good PC architectures, I'd be very interested to hear!