A “point cloud” is an important type of data structure for storing geometric shape data. Due to its irregular format, it’s often transformed into regular 3D voxel grids or collections of images before being used in deep learning applications, a step that makes the data unnecessarily large.
In our latest example (Soumik and myself), we present PointNet (from 2017) solves this problem by directly consuming point clouds, respecting the permutation-invariance property of the point data. Additionally, we’re working on a comprehensive repository on performing point cloud segmentation at scale with full TPU support.
