point-transformer-pytorch

Why we included this project

The Point Transformer attention layer from the 2020 paper, which set strong results on point cloud classification and segmentation, is packaged here as a single PyTorch module. You feed it point features and coordinates, and it returns updated features, with an optional k-nearest-neighbors setting to keep attention local on large point sets. It's a focused building block rather than a full pipeline, so it suits researchers and engineers who already have their own data loaders and downstream heads and just want a faithful reference implementation of the attention mechanism. The code is compact and readable, which makes it easy to adapt for custom experiments or to reproduce the paper's approach in a 3D vision project.