Introduction
Establishing point-to-point correspondences between 3D shapes is a fundamental task in computer vision and graphics, with applications ranging from texture transfer to statistical shape analysis. A dominant paradigm in recent years has been the functional maps framework. This approach simplifies the matching problem by operating in a reduced spectral domain, typically defined by the eigenfunctions of the Laplace-Beltrami operator. While highly effective for near-isometric full shapes, this framework encounters significant hurdles when applied to partial shape matching—the task of aligning a partial scan (e.g., a limb) to a complete reference model.
The Limitation of Functional Map Layers
Most deep learning-based shape matching pipelines incorporate a differentiable "functional map layer" (FM-layer). This layer typically solves a least-squares problem to estimate the transformation between spectral coefficients. This paper identifies a critical flaw in this approach for partial shapes: the least-squares calculation involves an integration over the entire domain. When parts of the shape are missing, this integration introduces unavoidable errors that scale linearly with the area of the missing regions. Consequently, as partiality increases, the estimated functional map becomes increasingly noisy, degrading the quality of the final point-to-point correspondence.
Proposed Methodology: Direct Feature Matching
To overcome the limitations of FM-layers, the authors propose an unsupervised pipeline that bypasses the intermediate spectral mapping step during the correspondence estimation.
1. Direct Correspondence via Softmax
Instead of estimating a functional map matrix $C$, the method computes a soft correspondence matrix $P$ directly from the cosine similarity of pointwise features. These features are extracted using DiffusionNet, a robust Siamese backbone. The authors demonstrate that using a softmax operator with a low temperature parameter $\tau$ is superior to the least-squares approach. While FM-layer errors are linear, the softmax operation exponentially suppresses noisy correlations from missing regions, leading to more robust matches.
2. Unsupervised Loss Functions
The network is trained without ground-truth correspondences using two primary objectives:
- Gromov Distance ($L_G$): This core loss ensures that the mapping preserves geodesic distances between pairs of points, promoting near-isometry.
- Low-Pass Filter Regularization ($L_{lpf}$): To enforce smoothness, the authors introduce a novel loss that assumes low-frequency eigenfunctions of the full shape can approximate those of the partial shape. This allows for spectral regularization without relying on an explicit functional map.
Experimental Evaluation
The method was rigorously tested on several challenging benchmarks, including SHREC’16 (CUTS and HOLES) and a newly introduced dataset, PFAUST. PFAUST includes two levels of difficulty: PFAUST-M (medium partiality) and PFAUST-H (high partiality with topological changes).
The results indicate that the proposed approach significantly outperforms previous unsupervised methods. On the SHREC’16 HOLES benchmark, it achieved a Mean Geodesic Error (MGE) of 10.2, surpassing the leading supervised method, DPFM (13.9), and the best unsupervised baseline (16.2).
Discussion and Conclusion
By identifying and removing the "bottleneck" of differentiable functional map layers, this work sets a new state-of-the-art for unsupervised partial shape matching. The key insight is that while spectral representations are powerful for regularization, they can be detrimental when used as an intermediate mapping layer for partial data.
Despite its success, some challenges remain. The current formulation is optimized for part-to-full matching and may struggle with part-to-part scenarios. Additionally, geodesic distances can be sensitive to large topological changes (the "wormhole" effect). Future work will explore scale-invariant metrics and more robust handling of varying topologies in real-world scans.
Summary updated