Variational Relational Point Completion Network for Robust 3D Classification.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Real-scanned 3D point clouds are often incomplete, hindering 3D modeling and perception.
• Existing methods generate coarse shapes lacking fine local details and overlook object structures.
• Deterministic mappings fail to capture the inherent uncertainty in point cloud completion.

Purpose of the Study:

• To propose a novel variational framework, Variational Relational point Completion network (VRCNet), for robust point cloud completion.
• To address limitations of existing methods by incorporating probabilistic modeling and relational enhancement.
• To introduce new multi-view partial point cloud datasets for comprehensive evaluation.

Main Methods:

• A dual-path architecture for probabilistic modeling using a point VAE for reconstruction and a generative path guided by the reconstruction distribution.
• Relational enhancement through point self-attention and selective kernel modules to refine local details.
• Development of multi-view partial point cloud datasets (MVP and MVP-40) with over 200,000 scans.

Main Results:

• VRCNet outperforms state-of-the-art methods on standard point cloud completion benchmarks.
• Demonstrated generalizability and robustness on real-world partial point cloud scans.
• Achieved robust 3D classification for partial point clouds, significantly increasing accuracy.

Conclusions:

• VRCNet offers a principled probabilistic approach to point cloud completion, improving both global structure and local details.
• The proposed relational enhancement effectively refines shape details by exploiting point feature relationships.
• VRCNet provides a robust solution for 3D geometric modeling and perception tasks involving incomplete point cloud data.