pointnet2

Why we included this project

PointNet++ is the implementation that accompanies the Stanford paper most point cloud work cites, and it remains a sensible place to start when your data is raw 3D points rather than images. It handles the two jobs most downstream projects need: classifying whole point sets and labeling individual points inside them, and the repo includes pretrained models, training scripts, and data processing utilities alongside the networks. Its main advance over the original PointNet is the architecture: it builds hierarchical features over local neighborhoods, much as convolutions do on images, and it copes with the uneven density real 3D sensors produce, which the earlier model did not. Anyone doing object recognition, part segmentation, or work on LiDAR and depth data will find this a solid baseline to measure newer approaches against.