TopologyFormer: structure transformer assisted topology reconstruction for point cloud completion

• Published in: Multimedia tools and applications Vol. 83; no. 26; pp. 68743 - 68771
• Main Authors: Jiang, Zhenwei, Gao, Chenqiang, Li, Pengcheng, Liu, Chuandong, Liu, Fangcen, Zhu, Lijie
• Format: Journal Article
• Language: English
• Published: New York Springer US 26-01-2024; Springer Nature B.V
• Subjects: Computer Communication Networks; Computer Science; Data Structures and Information Theory; Image reconstruction; Methods; Modules; Multimedia; Multimedia Information Systems; Neighborhoods; Representations; Sampling; Special Purpose and Application-Based Systems; Topology; Track 6: Computer Vision for Multimedia Applications; Point cloud; Encoder-decoder; Topology reconstruction; Point cloud completion
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-18136-9

Point cloud completion is a fundamental task to enhance the completeness and authenticity of point cloud data captured in the real world. Existing mainstream methods tend to obtain the complete point cloud by refining or up-sampling the coarse point cloud in a coarse-to-fine manner. However, these methods ingore the long-distance correlations between point clouds and obtain blurred topological structures, leading to potential challenges for downstream tasks. In this paper, we present a novel point cloud completion model, called TopologyFormer. After obtaining a coarse point cloud, our model restores the missing topology structure by directly reconstructing the coarse point cloud itself instead of relying on point displacements for refining or up-sampling the coarse point cloud. Specifically, the proposed method contains three main components, a Structure Completion Network, a topology reconstruction module and a feature enhancement module, the latter two of which compose a Topology-Aware Reconstruction Network. The structure completion network considers long-range correlations among different parts of the point cloud and emphasizes the global geometric structure of the incomplete point cloud, which completes the missing point cloud into an intermediate representation. The topology reconstruction module receives the intermediate representation from the structure completion network to recover the local shapes and even point distribution, aiming to generate a point cloud with a higher-quality topology. Furthermore, the feature enhancement module ensures more reliable topology information and allows further use of the global feature obtained by the structure completion network. Extensive experiments on public benchmarks including ShapeNet and real-scanned dataset KITTI demonstrate that our method is superior to the existing state-of-the-art methods.