MMCNet: Multi-Modal Point Cloud Completion Network Based on Self-Projected Views

Obtaining defect-free point cloud data is challenging due to performance constraints of acquisition devices and unavoidable occlusion, making point cloud data completion critical. In recent years, the rapid development of deep learning technology enabling increasingly abundant research on point cloud completion methods, with many showing excellent performance. In particular, multi-modal methods have attracted attention for effectively overcoming inherent data deficiencies. This study proposes a novel Multi-Modal Point Cloud Completion Network (MMCNet) based on self-projected view. The network uses multi-view fusion, dual-way fusion, and dual attention upsampling modules to effectively complete incomplete point clouds. The multi-view fusion module fuses image features from multiple perspectives, and the dual-way fusion module fuses point cloud and image features to realize the reconstruction of the rough point cloud. Then, the dual attention upsampling module refines the rough point cloud, focusing on missing areas while maintaining structural integrity, to recover high-quality complete point clouds. Experiments on multiple standard datasets validate the superiority of our proposed MMCNet over state-of-the-art completion networks.