Low-Level Graph Convolution Network for Point Cloud Processing

Hongyu Yan, Zhihong Wu, Li Lu

Abstract: Enlightened by the success of graph neural network, recent graph-based methods achieve impressive performance in point cloud processing. However, these methods usually employ high-level semantic edge features to update the central point feature. Less attention has been paid to the low-level geometric edge information that provides the geometric structure to enhance the model’s robustness for the coordinates and rigid transformation. To solve this drawback, we propose a novel graph convolution, named Low-level Graph Convolution (LGConv), for point cloud processing. Specifically, LGConv combines semantic and geometric edge features to update the central point feature. In addition, we propose a Divisible Attention Mechanism (DAM) to our LGConv to weigh the contributions of different neighbor nodes. With the proposed LGConv, we design our LGCNet for 3D shape classification and part segmentation. Extensive experiments demonstrate the superiority of LGCNet compared with previous graph convolution methods.