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

Area of Science:

• Computer Vision
• 3D Computer Graphics
• Machine Learning

Background:

• Point cloud completion from partial data is difficult due to the trade-off between global structure and local details.
• Severe incompleteness exacerbates challenges in 3D shape reconstruction.

Purpose of the Study:

• To propose a symmetry-guided progressive refinement network for robust point cloud completion.
• To address the challenge of balancing global consistency and local detail recovery in incomplete 3D shapes.

Main Methods:

• A Symmetry Graph Inference Network (SymGraphNet) learns structural correspondences using keypoint graphs without explicit symmetry plane estimation.
• A confidence-weighted Cross-Aware Decoder adaptively fuses features for balanced completion of visible and missing regions.
• A multi-stage residual refinement strategy progressively enhances geometric fidelity and point distribution.

Main Results:

• The proposed network achieves consistent improvements over state-of-the-art methods on PCN, MVP, and KITTI datasets.
• Demonstrated effectiveness in both synthetic and real-world incomplete point cloud scenarios.
• Successfully balances the recovery of global structure and fine-grained local details.

Conclusions:

• The symmetry-guided progressive refinement network offers a robust solution for point cloud completion under severe incompleteness.
• The method effectively learns structural correspondences and progressively refines shapes for improved 3D reconstruction.
• This approach advances the state-of-the-art in 3D shape completion from partial observations.