(A) Part-based approach for real-world point cloud completion

Abstract

Real-world point clouds often suffer from incompleteness caused by self-occlusion, occlusion by other objects, sensor noise, and errors in the acquisition process. Precise three-dimensional object recognition requires the reconstruction of these incomplete point clouds to their original shapes. However, applying models trained on synthetic data directly to real point clouds leads to suboptimal performance due to inherent differences between the domains. This study addresses this challenge by aiming to reduce domain gaps through the utilization of part attributes shared by objects within the same category. The part-based approach proposed here simplifies the complexity of reconstruction, as parts typically exhibit simpler topologies compared to the overall shape. Furthermore, by learning relationships between parts, the method predicts missing parts based on the existing ones. The advantage of this approach lies in the fact that objects within the same category share attributes of the same parts, regardless of the domain, contributing to the reduction of domain gaps. To effectively leverage the features of parts, this study introduces a part-based framework and a part discriminator. The part-based framework is employed to reconstruct each part individually, which is then assembled to form the complete point cloud. The part discriminator aligns the part features between synthetic and real-world point clouds, effectively mitigating domain gaps. Extensive experiments demonstrate that the proposed method surpasses state-of-the-art approaches in handling real-world point clouds.