M-GCN: Multi-scale Graph Convolutional Network for 3D Point Cloud Classification

Jing Hu, Xincheng Wang, Ziheng Liao, Tingsong Xiao

Abstract: Point clouds are among the popular geometric representations for 3D vision applications. 3D point clouds are irregular and flexible, thus processing and summarizing information over these unordered data points are very challenging. Although a number of previous works attempt to analyze point clouds and achieve promising performances, their performances still lack efficient topological information. In this paper, we propose a novel multi-scale graph convolutional network (M-GCN), which is designed to extract local geometric features based on multi-scale feature fusion. The extracted local topological information across scales can enrich the representation power of point clouds more effectively. Experiments on ModelNet40 show that local graph neural networks built with various scale point features and edge features can achieve state-of-the-art performance on challenging classification benchmarks of 3D point clouds.