[Paper Review] Kernel Methods on the Riemannian Manifold of Symmetric Positive Definite Matrices

Sadeep Jayasumana, Richard Hartley+3 more|arXiv (Cornell University)|Dec 12, 2014

Medical Image Segmentation Techniques4 citations

TL;DR

This paper proposes a family of provably positive definite kernels for symmetric positive definite (SPD) matrices by leveraging Riemannian geometry, specifically using geodesic distances derived from various metrics on the SPD manifold. By embedding SPD matrices into a reproducing kernel Hilbert space (RKHS) via these kernels, the authors enable the direct application of Euclidean kernel methods—such as SVM and kernel PCA—on the manifold, achieving superior performance over tangent space approximations and standard Euclidean kernels in pedestrian detection, object categorization, texture analysis, and medical image segmentation tasks.

ABSTRACT

Symmetric Positive Definite (SPD) matrices have become popular to encode image information. Accounting for the geometry of the Riemannian manifold of SPD matrices has proven key to the success of many algorithms. However, most existing methods only approximate the true shape of the manifold locally by its tangent plane. In this paper, inspired by kernel methods, we propose to map SPD matrices to a high dimensional Hilbert space where Euclidean geometry applies. To encode the geometry of the manifold in the mapping, we introduce a family of provably positive definite kernels on the Riemannian manifold of SPD matrices. These kernels are derived from the Gaussian ker- nel, but exploit different metrics on the manifold. This lets us extend kernel-based algorithms developed for Euclidean spaces, such as SVM and kernel PCA, to the Riemannian manifold of SPD matrices. We demonstrate the benefits of our approach on the problems of pedestrian detection, ob- ject categorization, texture analysis, 2D motion segmentation and Diffusion Tensor Imaging (DTI) segmentation.

Motivation & Objective

• To address the limitation of existing manifold-based algorithms that rely on local tangent space approximations, which distort true Riemannian geometry and incur high computational cost.
• To develop a theoretically grounded framework for constructing positive definite kernels on non-Euclidean manifolds, specifically the Riemannian manifold of SPD matrices.
• To enable the use of powerful kernel-based algorithms—such as SVM, kernel PCA, and kernel k-means—on SPD matrix data by embedding them into a high-dimensional Hilbert space with geometry-aware kernels.
• To demonstrate that geometry-aware kernels outperform both standard Euclidean kernels and tangent space-based methods across multiple vision tasks.

Proposed method

• Derives a family of positive definite Gaussian kernels on the SPD manifold by replacing the Euclidean distance in the standard Gaussian kernel with geodesic distances from valid Riemannian metrics.
• Establishes a theoretical framework to analyze the positive definiteness of Gaussian kernels based on distance functions, proving that certain Riemannian metrics on Sym+d yield valid positive definite kernels.
• Utilizes the log-Euclidean metric as a key example, which provides a true geodesic distance and enables efficient computation via matrix logarithms and exponentials.
• Applies the proposed kernels to extend standard kernel algorithms—SVM, kernel PCA, kernel k-means—to the Riemannian manifold of SPD matrices, preserving the intrinsic geometry of the data.
• Employs the kernel trick to implicitly map SPD matrices into a high-dimensional RKHS where Euclidean algorithms can be applied without explicit computation.
• Validates the framework using multiple computer vision tasks, including pedestrian detection, object categorization, texture recognition, and image segmentation in DTI and motion analysis.

Experimental results

Research questions

• RQ1 Can Gaussian kernels defined using geodesic distances on the SPD manifold be provably positive definite?
• RQ2 How can kernel-based algorithms like SVM and kernel PCA be generalized to the Riemannian manifold of SPD matrices?
• RQ3 Does using geometry-aware kernels on the SPD manifold lead to improved performance compared to Euclidean kernels and tangent space approximations?
• RQ4 Which Riemannian metrics on Sym+d yield valid positive definite kernels when used in Gaussian kernel formulations?

Key findings

• The Riemannian kernel-based kernel PCA achieved 96.40% accuracy on the Brodatz texture dataset with l=12, significantly outperforming the Euclidean kernel’s 91.89% under the same conditions.
• Kernel k-means with the Riemannian kernel successfully segmented the corpus callosum in a real DTI brain image, producing cleaner results than baseline methods that used dimensionality reduction via LLE, LE, or HLLE.
• On the Hamburg Taxi motion segmentation dataset, the proposed Riemannian kernel k-means produced more accurate and less noisy segmentations than the baseline methods that first reduced dimensionality before clustering.
• The proposed method avoids iterative logarithmic and exponential maps, reducing computational cost compared to tangent space-based approaches while preserving global geometric structure.
• The framework is generalizable to other manifolds, provided their metrics induce negative definite squared distances, suggesting broader applicability beyond SPD matrices.