PoinTr-PM: Diverse Point Cloud Completion with Geometry-Aware Transformers and Point Moving

Point Cloud are widely utilized data in addressing 3D environmental perception challenges. During the actual data collection process, the acquired raw point cloud data is often incomplete due to limitations such as sensor resolution, occlusion, relative object motion, and reflections. This incompleteness significantly affects the performance of downstream tasks based on point clouds. Therefore, recovering complete point cloud data from incomplete point cloud data has been an important and challenging task. PoinTr tackles the issue of recovering complete point cloud data from incomplete data by employing a Transformer network. However, its completion results often exhibit considerable noise. To address this issue, we propose an approach named PoinTr-PM, which integrates the concept of point movement with PoinTr. Experimental results indicate that our proposed method, PoinTr-PM, partially mitigates the noise issues present in PoinTr and, in certain indicators, outperforms PoinTr.