Point Cloud Completion via Relative Point Position Encoding and Regional Attention

Abstract

The global feature encoding and surface detail refinement are two critical components for point-based point cloud completion methods. However, existing methods typically use max pooling to hard integrate the neighbouring features, resulting in that the global feature can not well encode the majority of point position information. Moreover, as the important factor of refinement, the position displacement is not well represented and suffers from the information loss of structure details on local and non-local regions. Thus we propose a novel regional attention-based Siamese auto-encoder network architecture, by which the majority of relative point position information is well encoded in the global feature. Then a low order local attention and a high order non-local attention are presented to search the contributive local and non-local features for regressing the position displacements of shape surface. Quantitative and qualitative experiments on PCN, Completion3D, MVP, ShapeNet-55/34, and KITTI datasets show that the proposed method achieves competitive results compared with existing state-of-the-art completion methods.