CCT-Net: Contour Collaborative Transformer for Point Cloud Completion

Point cloud completion aims to recover missing regions from the partial point cloud collected by sensors. Existing methods typically extract local features from the input point cloud to recover its regional details. However, they ignore the overall shape and the edge information, which in contrast are the advantages of point cloud contour. In this paper, a new Contour Collaborative Transformer network (CCT-Net) is proposed to achieve the complementary advantages of contour features and local features. Specifically, the input point cloud is processed by two different feature extractors to capture its local features and fine-grained contour features characterizing the overall shape and edge information. In addition, a skeleton generator is proposed to facilitate the complementarity of both contour and local features by aligning and enhancing them for generating a relatively complete shape skeleton. After that, a point generator is introduced to refine the shape skeleton into a complete point cloud with rich details using the contour features. Extensive experiments on PCN, ShapeNet-55, and MVP datasets demonstrate that our CCT-Net outperforms the state-of-the-art methods.