SkeletonFormer: Point Cloud Completion with Dynamic Selective Skeleton Points

Point cloud completion aims at recovering the complete point cloud from an incomplete input. A general scheme is to generate a group of coarse points first that generalize the global shape, and then reconstruct dense point cloud by upsampling operation. In this paper, we propose a novel point cloud completion network, SkeletonFormer, to tackle two critical challenges: fully utilizing the information from the point cloud with various incompleteness degree and recovering high-quality geometric structures. To increase the universality of our model to diverse input, we propose a score mechanism to dynamically select proper skeleton points that can adapt to various degree of deficiency. To improve the perception of the target object, we use self-projected depth images as an augmented modality representation to observe the input. A modality unification module is designed to fuse the depth image feature and the point cloud feature, and it can alleviate the intrinsic differences among multi-modal information. The fused feature is used to assist the prediction of the skeleton points that represent the holistic complete object. Furthermore, by fully leveraging local geometric information, we design a novel and effective DeconvNet to reconstruct fine-grained patterns around the skeleton points. Extensive experiments demonstrate that our SkeletonFormer surpasses existing works by a large margin and achieves state-of-the-art performance on various benchmarks.