Using deep learning in semantic classification for point cloud data

Abstract

Point cloud is an important 3D data structure, but its irregular format brings great challenges to deep learning. The advent of PointNet makes it possible to process irregular point cloud data by neural networks directly. As an extension of PointNet, PointNet++ can extract local features, which makes it perform better than PointNet in processing point cloud data. But in practice, it is common that the density of a point set usually varies with the location, which makes the computation overhead of PointNet++ very heavy. To deal with it, we propose an octree grouping-based network structure for PointNet++, named Octree-Grouping-PointNet++ (OG-PointNet++). It determines the point density by constructing an unbalanced octree for the point cloud, and groups point according to the point density. These point groups are assigned to different layers according to their density, and the local feature of each group is extracted by PointNet++. The global feature is obtained from the last abstract layer and used for classification and segmentation. The experiments show its competitive performance in many 3D tasks, such as object classification and semantic segmentation.