Non-Rigid Point Cloud Registration via Anisotropic Hybrid Field Harmonization.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Computational Geometry

Background:

• Point cloud registration algorithms face challenges with simultaneous deformations and occlusions.
• Existing methods suffer from inaccurate manifold modeling and suboptimal optimization for fragmented structures.

Purpose of the Study:

• To introduce AniSym-Net, a novel framework for non-rigid point cloud registration.
• To address challenges in near-isometric deformation registration under occlusions.

Main Methods:

• Developed a novel anisotropic hybrid shape-motion deformation field to encode topological and geometric information.
• Utilized holonomic constraints from symplectic structure modeling and motion-conditional cross-attention.
• Optimized loop closures of cotangent bundles within a symplectic manifold framework.

Main Results:

• AniSym-Net significantly mitigates registration errors and occlusions by harmonizing correspondences.
• Demonstrated superior performance over state-of-the-art methods on five popular benchmarks (CAPE, DT4D, SAPIEN, FAUST, DeepDeform).

Conclusions:

• AniSym-Net offers a robust solution for non-rigid point cloud registration with deformations and occlusions.
• The proposed framework advances the field by effectively modeling complex shape dynamics and improving registration accuracy.