charlesq34/pointnet2
A deep learning architecture that learns hierarchical features from 3D point clouds for classification and segmentation tasks.

PointNet++ extends the original PointNet by respecting spatial locality in point sets, enabling hierarchical feature learning across multiple scales. It processes 3D point clouds directly without requiring mesh or voxel conversion, addressing non-uniform density challenges inherent in real-world sensor data. The architecture implements set abstraction layers that progressively downsample point sets and capture local geometric patterns, similar to how convnets operate on image grids.