Learning semantic keypoints for diverse point cloud completion

Raw point clouds collected from real-world scenes are sparse, incomplete and noisy, posing significant challenges for their integration into automation workflows in construction. Thus, completing plausible and fine-grained point clouds is a critical prerequisite for downstream applications. Current methods primarily focus on learning patch-level features and modeling their relationships for inferring complete object shapes. However, the significant disparity between real-world scenarios and clean synthetic datasets limits their representation ability of local structures, especially when facing noises and irregular missing patterns. This paper proposes a semantic keypoint guided completion network (SKPNet) to enhance the generalization ability of point cloud completion in diverse construction scenarios in a semantic-guided manner. The key insight is to build a connection between the object geometric structure and its global semantic feature, which is more robust to point-level disruptions. Accordingly, a semantic keypoint generation module is developed to learn representative keypoints based on the global semantic vector encoded from the input points. These keypoints then serve as the control points for searching the neighboring point-level features with rich local pattern information, simultaneously filtering out the noises during the process. By progressively incorporating multi-scale point-level features, this paper gradually refines and upsamples the keypoints to the final fine-grained completion. Comprehensive experiments conducted on both synthetic and real-world datasets demonstrate the competitive and robust performance of SKPNet in completing high-quality shapes.