torch-points3d

Why we included this project

Point clouds don't map cleanly onto standard neural network layers, and this framework is built to spare teams working in 3D that pain. It groups a set of familiar architectures (PointNet, PointNet++, KPConv, RandLA-Net, VoteNet, and others) behind one consistent PyTorch API, so you can switch models without rebuilding your training and evaluation loop. Semantic segmentation, classification, registration, object detection, and panoptic segmentation each have their own dataset and model organization, which makes comparing runs on S3DIS and ScanNet much less fiddly. Because it sits on PyTorch Geometric and uses Hydra for configuration, reproducibility comes with the design rather than being added later. That is most useful to research groups and ML engineers who want to try several point-cloud approaches quickly or reproduce published results without reimplementing each network from scratch.