Reference Convolutional Networks for 3D Deep Point Signature Learning

Learning geometric features directly from point clouds has proven efficient and adaptive in many tasks. Although containing comprehensive information, raw point cloud inputs only provide the geometric information of a 3D object or scene. This paper presents Ref-CNN, a convolutional network with multiple references for point signature learning using point clouds. Ref-CNN learns point signatures without increasing computation, drawing from multiple reference planes for each point cloud. Specifically, this paper defines reference planes for each point set to extract point signatures from multiple perspectives. We then formulate a method to expand the dimension of each point vector, enhancing its spatial location information in multiple reference systems. Additionally, we design a permutation invariant network structure to actively learn local and global features from the expanded data structure of the point cloud. Experiments on popular shape recognition benchmarks demonstrate Ref-CNN's performance.