A Two-Stage Transformer-Based Point Cloud Upsampling

Point cloud upsampling is used to densify the sparse set of points collected by 3D sensors, which is widely used in the field of robotics. In this paper, a two-stage method is used to generate upsampling point cloud. In the first stage, the rough upsampling point cloud is generated, and in the second stage, the rough point cloud is refined to obtain high-quality point cloud. The generation of high quality point cloud heavily relies on feature extractors. In the first stage, we introduce a transformer model to improve the feature extraction effect, which helps to fully extract the features of different location points. A Refinement Unit is proposed in the second stage to improve the rough upsampling point cloud generated in the first stage. The Refinement Unit is based on another transformer model in which the relative position coding function improves the coordinates of the deviation points generated in the first stage. We evaluate our approach on synthetic data sets. Experimental results show that the performance of this method is better than other methods.