Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),
September 2021 · Prague, Czech Republic
Inferring a complete 3D geometry given an incomplete point cloud is essential in many vision and robotics applications. Previous work mainly relies on a global feature extracted by a Multi-layer Perceptron (MLP) for predicting the shape geometry. This suffers from a loss of structural details, as its point generator fails to capture the detailed topology and structure of point clouds using only the global features. The irregular nature of point clouds makes this task more challenging. This paper presents a novel method for shape completion to address this problem. The Transformer structure is currently a standard approach for natural language processing tasks and its inherent nature of permutation invariance makes it well suited for learning point clouds. Furthermore, the Transformer's attention mechanism can effectively capture the local context within a point cloud and efficiently exploit its incomplete local structure details. A morphing-atlas-based point generation network further fully utilizes the extracted point Transformer feature to predict the missing region using charts defined on the shape. Shape completion is achieved via the concatenation of all predicting charts on the surface. Extensive experiments on the Completion3D and KITTI data sets demonstrate that the proposed PCTMA-Net outperforms the state-of-the-art shape completion approaches and has a 10% relative improvement over the next best-performing method.