Abstract
Non-rigid shape registration, the process of finding correspondences and interpolations between 3D shapes, remains a core challenge in computer vision. This paper introduces a novel framework that addresses these tasks simultaneously by modeling deformation as a volume-preserving flow field. By enforcing a zero-divergence constraint and utilizing a coarse-to-fine spectral basis, the method produces physically plausible, resolution-independent results.
Introduction
Finding the "right" way to morph one 3D shape into another is a fundamental problem in computer vision. Traditional methods often decouple shape correspondence (matching points) from interpolation (generating intermediate frames). Simple techniques like linear interpolation frequently fail to respect physical properties, leading to unrealistic "squeezing" or volume loss during large rotations. Moreover, many existing approaches are computationally intensive and tied to specific mesh resolutions, hindering scalability to high-detail scans.
Methodology
The proposed approach models deformation as a continuous flow field $v$ in 3D space, built on three key pillars:
- Divergence-Free Constraint: Inspired by fluid dynamics, the velocity field is required to have zero divergence ($\nabla \cdot v = 0$). This ensures volume preservation throughout the deformation. This is mathematically enforced by representing the velocity field as the curl of a potential field $\Phi$ ($v = \nabla \times \Phi$).
- Karhunen-Loève Basis: To avoid the high cost of volumetric discretization, the potential field $\Phi$ is represented using a truncated set of eigenfunctions of the Laplacian operator. This provides a natural coarse-to-fine hierarchy for capturing both global and fine-grained motion.
- Joint Optimization (EM Framework): An Expectation-Maximization (EM) approach solves for deformation and correspondence concurrently. The E-step updates soft correspondences using Euclidean distance and SHOT descriptors, while the M-step optimizes the deformation field coefficients via a Gauss-Newton method.
Final paths are computed using a 2nd-order Runge-Kutta integration scheme, providing higher accuracy in volume preservation over time compared to simpler Euler integration.
Datasets and Evaluation
The framework was evaluated across several benchmark datasets:
- TOSCA: 76 meshes across 8 classes for intra-class correspondence testing.
- FAUST: High-resolution human scans (200k+ vertices) often containing topological noise.
- KIDS & AIM@SHAPE: Used for qualitative demonstrations of complex morphs and extrapolations, specifically the Armadillo model.
Results and Key Insights
The method achieves state-of-the-art accuracy on the Princeton benchmark (TOSCA), matching or exceeding established techniques like Functional Maps and Coherent Point Drift. Key advantages include:
- Physical Realism: The divergence-free constraint acts as a powerful regularizer, preventing "collapsing" artifacts and producing natural-looking morphs for articulated objects.
- Resolution Independence: The continuous basis allows optimization on downsampled meshes (e.g., 3,000 vertices) with the results applied seamlessly to high-resolution scans (200k+ vertices) with negligible extra cost.
- Extrapolation: The time-independent nature of the velocity field allows for plausible pose extrapolation by simply extending the integration time past the target shape.
Discussion
While effective, the method has limitations. The "global" volume preservation constraint can be too rigid when matching individuals with different body types (e.g., thin vs. muscular). Additionally, the framework struggles with topological changes, such as hands touching and separating, as the volume preservation applies to the empty space between them as well. Future work will explore locally volume-preserving and time-dependent deformation fields to handle more complex, multi-stage motions.
Summary updated