Structure-Preserving Point Cloud Completion with Symmetry-Guided Progressive Refinement

Point cloud completion from partial observations remains challenging due to the trade-off between preserving global structural consistency and recovering fine-grained local details, especially under severe incompleteness. We propose a symmetry-guided progressive refinement network to address this problem by learning flexible structural correspondences and progressively refining incomplete shapes. First, a Symmetry Graph Inference Network (SymGraphNet) constructs a feature-space graph over sampled keypoints and predicts symmetry-guided structural counterparts for robust coarse shape recovery, without explicitly estimating a rigid symmetry plane or axis. Second, a confidence-weighted Cross-Aware Decoder adaptively fuses partial-observation features and symmetry-guided features to balance visible-region fidelity and missing-region completion. Third, a multi-stage residual refinement strategy progressively improves geometric fidelity, local continuity, and point distribution uniformity. Experiments on PCN, MVP, and KITTI datasets demonstrate consistent improvements over representative state-of-the-art methods under both synthetic and real-world incomplete point cloud settings.