A Point Cloud Upsampling Adversarial Network Based on Residual Multi-Scale Off-Set Attention

Abstract

Due to the limitation of the working principle of 3D scanning equipment, the point cloud obtained by 3D scanning is usually sparse and unevenly distributed. In this paper, we propose a new Generative Adversarial Network(GAN) for point cloud upsampling, which is extended from PU-GAN. Its core architecture is to replace the traditional Self-Attention (SA) module with the implicit Laplacian Off-Set Attention(OA) module, and adjacency features are aggregated using the Multi-Scale Off-Set Attention(MSOA) module, which adaptively adjusts the receptive field to learn various structural features. Finally, Residual links were added to form our Residual Multi-Scale Off-Set Attention (RMSOA) module, which utilized multi-scale structural relationships to generate finer details. A large number of experiments show that the performance of our method is superior to the existing methods, and our model has high robustness.