Point clouds size

[danlapko]

Hengshuang et al mention in their work "we apply selfattention locally, which enables scalability to large scenes with millions of points". But this implementation could be hardly trained for num_point>8k (nvidia rtx 3090).
Any suggestions on how to train/apply this implementation for large point clouds?

[qq456cvb]

You may borrow some ideas from PointNet++, which splits the whole area into a set of chunks, e.g., 1m x 1m x 1m. Then within each chunk, use the PointTransformer to predict the label for each point. You may find more details about this in the code provided by PointNet++ (https://github.com/charlesq34/pointnet2/blob/master/scannet/scannet_dataset.py).

[danlapko]

Got it, thank you!

[qq456cvb]

This repo is a compact, educational re-implementation that uses global FPS and kNN, so it isn't built for million-point scenes (memory blows up on the pairwise distances). The scalability in the paper comes from applying attention locally inside a block or sliding-window pipeline, like their S3DIS setup. To go large you'd process the scene in chunks and use an efficient CUDA kNN, which is out of scope here.