FPTNet: Full Point Transformer Network for Point Cloud Completion

In this paper, we propose a novel point transformer network (FPTNet) for point cloud completion. Firstly, we exploit the local details as well as the long term relationships from incomplete point shapes via residual point transformer blocks. Secondly, we realize the deterministic mapping learning is a challenging task as point completion is a many-to-one problem. To address this, the shape memory layer is designed to store general shape features. The network infers complete point shapes from both incomplete clouds and shape memory features. Thirdly, the recurrent learning strategy is proposed to gradually refine the complete shape. Comprehensive experiments demonstrate that our method outperforms state-of-the-art methods on PCN and Completion3D benchmarks.