Hybrid Geometric Primitive Representations for Large Scale Point Clouds

Arriola, M.*, Huang, X.*, Wang, Y., Guizilini, V.C., Ambrus, R.A., Solomon, J.

Point clouds, collected by LiDAR scanners and other 3D sensing devices, are commonly used by deep learning models to perceive the physical world. However, they present several critical challenges for machine learning thanks to their irregularity and sparsity. We propose a heterogeneous graph neural network architecture for point cloud data that distinguishes between two node types: isolated points and geometric primitives like line segments, planar patches and volumetric boxes. By recognizing the existence of geometric primitives, our architecture achieves high efficiency and performance on several key point cloud processing tasks.