Li Jiaxin Chen Ben M. Lee Gim Hee

Abstract

This paper presents SO-Net, a permutation invariant architecture for deeplearning with orderless point clouds. The SO-Net models the spatialdistribution of point cloud by building a Self-Organizing Map (SOM). Based onthe SOM, SO-Net performs hierarchical feature extraction on individual pointsand SOM nodes, and ultimately represents the input point cloud by a singlefeature vector. The receptive field of the network can be systematicallyadjusted by conducting point-to-node k nearest neighbor search. In recognitiontasks such as point cloud reconstruction, classification, object partsegmentation and shape retrieval, our proposed network demonstrates performancethat is similar with or better than state-of-the-art approaches. In addition,the training speed is significantly faster than existing point cloudrecognition networks because of the parallelizability and simplicity of theproposed architecture. Our code is available at the project website.https://github.com/lijx10/SO-Net