Multiway Non-Rigid Point Cloud Registration via Learned Functional Map Synchronization.

Area of Science:

• Computer Vision
• Computational Geometry
• 3D Shape Analysis

Background:

• Non-rigid shape registration is crucial for applications like computer animation and 3D digitization.
• Existing methods struggle with real-world data, including noisy scans and occlusions.
• A robust and flexible framework for multi-shape alignment is lacking.

Purpose of the Study:

• To introduce SyNoRiM, a novel method for the joint registration of multiple non-rigid shapes.
• To address challenges posed by noisy scans and occlusions in shape correspondence.
• To develop a unified framework for both non-rigid and multi-body shape registration.

Main Methods:

• Utilizes functional maps to parameterize pairwise correspondences between point clouds.
• Simultaneously learns non-orthogonal basis functions for deformation regularization and occlusion handling.
• Employs a novel optimization formulation to synchronize pairwise functional maps into a cycle-consistent whole.

Main Results:

• Achieves state-of-the-art performance in registration accuracy through extensive experiments.
• Demonstrates flexibility by handling both non-rigid and multi-body registration within a unified framework.
• Avoids computationally expensive point-wise permutation optimization by using basis function maps.

Conclusions:

• SyNoRiM provides a robust, flexible, and efficient solution for multi-shape registration.
• The method effectively handles occlusions and noisy data in non-rigid shape analysis.
• SyNoRiM advances the field of 3D shape correspondence and alignment.