Point Cloud Upsampling Network Incorporating Dynamic Graph Convolution and Multi-Head Attention

Abstract

To address the problems that graph convolution uses a fixed graph structure, fails to capture dynamic or changing graph structure information, and is prone to bias by employing the same attention. A point-cloud upsampling network (DGCMSA-PU) incorporating Dynamic Graph Convolutional (DGCNN) and Multi-head Self-Attention (MHSA) is designed. DGCNN is utilised for up-sampling and a MHSA mechanism is incorporated to simultaneously fuse information from different attention heads. The edge relationships between nodes in the graph data are captured by edge convolution (EdgeConv), and the graph structure is dynamically constructed based on the relationships between nodes. Then the features of the point cloud are extracted by the three attention heads with different weights and different foci. Finally, an up-down-up structure is used to extend the features and reconstruct the 3D coordinates of the output point cloud. The superiority of DGCMSA-PU in the up-sampling task is verified through experiments comparing it with existing up-sampling networks, and the robustness of the network to noise and varying number of input point clouds, as well as the important role of the Multi Headed Attention module in the performance improvement of the network, are analysed through robustness and ablation experiments.