Abstract
This article presents a new distance for measuring shape dissimilarity between objects. Recent publications introduced the use of eigenvalues of the Laplace operator as compact shape descriptors. Here, we revisit the eigenvalues to define a proper distance, called Weighted Spectral Distance (WESD), for quantifying shape dissimilarity. The definition of WESD is derived through analysing the heat-trace. This analysis provides the proposed distance an intuitive meaning and mathematically links it to the intrinsic geometry of objects. We analyse the resulting distance definition, present and prove its important theoretical properties. Some of these properties include: i) WESD is defined over the entire sequence of eigenvalues yet it is guaranteed to converge, ii) it is a pseudometric, iii) it is accurately approximated with a finite number of eigenvalues, and iv) it can be mapped to the [0,1) interval. Lastly, experiments conducted on synthetic and real objects are presented. These experiments highlight the practical benefits of WESD for applications in vision and medical image analysis.
Introduction
Quantifying the dissimilarity between shapes is a cornerstone of computer vision, pattern recognition, and medical imaging. A robust shape metric must be invariant to isometric transformations—such as rotation, translation, and changes in pose—while remaining sensitive to meaningful geometric differences. A popular framework for achieving this involves the spectrum (eigenvalues) of the Laplace-Beltrami operator. These eigenvalues act as a "fingerprint" of the shape, analogous to the resonant frequencies of a vibrating membrane.
However, existing spectral descriptors, such as Shape-DNA, face significant theoretical and practical hurdles. Because eigenvalues grow toward infinity, standard Euclidean distances between eigenvalue sequences diverge. Consequently, researchers are often forced to truncate the sequence at an arbitrary number of modes ($N$). This makes the resulting distance highly sensitive to the choice of $N$ and causes higher-frequency modes—which are often dominated by noise—to disproportionately influence the metric.
The Weighted Spectral Distance (WESD)
To address these limitations, we introduce the Weighted Spectral Distance (WESD). The core innovation lies in a mathematically grounded weighting scheme derived from the heat-trace of the object. For two shapes $\Omega_\lambda$ and $\Omega_\xi$ with eigenvalue sequences $\{\lambda_n\}$ and $\{\xi_n\}$, the WESD is defined as:
$$\rho(\Omega_\lambda, \Omega_\xi) = \left[ \sum_{n=1}^{\infty} \left( \frac{|\lambda_n - \xi_n|}{\lambda_n \xi_n} \right)^p \right]^{1/p}$$
The denominator $\lambda_n \xi_n$ serves as a natural regularizer. By weighting the differences by the inverse of the eigenvalues, the contribution of higher-frequency modes is diminished. This ensures that the infinite sum converges to a finite value for $p > d/2$ (where $d$ is the dimensionality of the object).
Spectral Shape Representation
The relationship between the spectrum and the geometry is best understood through the eigenfunctions. While eigenvalues provide the global distance metric, the eigenfunctions represent the spatial distribution of these "vibrational" modes across the object's surface.
As illustrated in Figure 1, lower-order eigenfunctions capture global structure, while higher-order modes correspond to increasingly localized and complex geometric details. WESD effectively integrates these multi-scale features into a single, stable distance measure.
Theoretical Properties and Advantages
WESD offers several key advantages over traditional spectral methods:
- Convergence and Stability: Unlike previous methods, WESD is stable with respect to the number of eigenvalues used. The error introduced by using a finite number of modes ($N$) has a theoretical upper bound that vanishes as $N \to \infty$.
- Multi-scale Sensitivity: The parameter $p$ allows for tuning the metric's sensitivity. Higher values of $p$ emphasize global, coarse features, while lower values allow for finer discrimination of small-scale geometric variations.
- Normalization: The metric can be mapped to a $[0, 1)$ interval (nWESD), facilitating easier comparison across different datasets and integration into machine learning pipelines.
- Physical Grounding: By deriving the distance from the heat-trace, WESD is rooted in the physics of diffusion, providing a more principled approach than heuristic feature-vector comparisons.
Experimental Evaluation
The effectiveness of WESD was validated across several diverse benchmarks:
- Shape Retrieval: On the SHREC 2011 benchmark of non-rigid 3D objects, WESD achieved state-of-the-art retrieval accuracy while demonstrating superior stability across different parameter settings compared to Shape-DNA.
- Medical Image Analysis: In studies of human brain MRI (LPBA40), WESD successfully clustered subcortical structures like the hippocampus and putamen, maintaining distinct clusters even in the presence of manual segmentation artifacts.
- Dynamic Cardiac Analysis: Using 4D cardiac MRI, the normalized WESD (nWESD) accurately tracked the deformation of the left ventricle throughout the cardiac cycle, showing high symmetry between the contraction (systole) and relaxation (diastole) phases.
Conclusion
WESD provides a mathematically rigorous and practically robust framework for shape comparison. By solving the divergence and instability issues inherent in previous spectral methods, it enables more reliable shape analysis in complex domains such as medical imaging and 3D computer vision. Future work will explore the impact of different boundary conditions and the application of WESD to even larger-scale geometric datasets.
Summary updated