A new self-augment CNN for 3D point cloud classification and segmentation

Xinhong Meng, Xinyu Lu, Hailiang Ye, Bing Yang, Feilong Cao

Abstract: Point cloud classification and segmentation are challenging tasks due to the irregular structures, especially when there is translation variance in the point clouds. To overcome this barrier, this paper proposes a self-augment convolutional neural network (SACNN), which can not only extract more discriminative features from the points cloud but also alleviate the translation variance problem. Specifically, we first represent the point cloud through the dynamic graph, with the intention to keep the number of point clouds during the feature learning to avoid information loss. Benefiting from the dynamic graph, the global and local features of point clouds can be learned. Then, to reduce the translation variance in the dynamic graphs, a self-augment convolution (SAConv) module is designed to make points align their coordinates based on learned features. Finally, the local mixed aggregation module is proposed to combine the overview and the detailed descriptor of the neighbors. Experiments on several standard benchmarks verify the superiority of the SACNN over state-of-the-art methods in both 3D point cloud classification and segmentation tasks.