A Point Cloud Completion Network Based on Multi-Channel Resolution Transformer

In the process of point cloud data collection, the obtained point cloud is usually sparse, noisy, and incomplete. Therefore, restoring complete, high-precision, and high-fidelity point cloud from partial point cloud is an essential task in many visual and robotics applications. This article proposes a point cloud completion method based on multi-channel resolution Transformer (MCREPT), and designs a multi-channel resolution feature extractor (MCRE) that can effectively capture the coordinate space and feature space of local center point clouds. This will introduce rich point cloud feature at low, medium, and high levels of different receptive fields, effectively solving the problem of lack induction bias in Transformer and insufficient richness in point cloud feature hierarchy. In the encoder decoder of Transformer, this paper proposes a geometric detail perceptron based on MCR to enhance the extraction of local shape detail features, solving the problem of missing induction bias in Transformer. A large number of experimental results from PCN and KITTI have verified that MCREPT generates more accurate and uniform geometric shapes than existing Transformer based point cloud completion methods.