Point Cloud Upsampling Network Using Spatially Biased Adaptive Graph Convolution

Abstract

Point cloud data plays a vital role in representing the three-dimensional (3D) world and finds extensive use in autonomous driving and three-dimensional cultural heritage displays. However, technical and cost constraints often lead to acquired point cloud data exhibiting density disparity, making it unsuitable for direct use in downstream tasks. To address this issue, this paper proposes a novel network called PU-SpaAGCN, using spatially biased adaptive graph convolution. PU-SpaAGCN overcomes the limitations associated with traditional graph convolution, such as the presence of isotropic edge features, by effectively capturing precise point cloud edges and local features. Additionally, the network introduces a spatial bias term to the graph convolution, facilitating the extraction of global features. The fusion of local and global features enables the network to generate higher-quality dense point clouds. Furthermore, unlike previous point cloud upsampling networks, the network introduces a feature space interpolation module, enhancing both network performance and interpretability. Experimental evaluations conducted on the PU1K dataset demonstrate that the dense point cloud produced by PU-SpaAGCN exhibits significant improvements in terms of point cloud uniformity and realism.