Kernel Methods on Riemannian Manifolds with Gaussian RBF Kernels.

Area of Science:

• Computer Vision
• Machine Learning
• Differential Geometry

Background:

• Computer vision often involves data on Riemannian manifolds with non-Euclidean geometry.
• Standard Euclidean algorithms perform poorly on such manifold-valued data.

Purpose of the Study:

• To develop positive definite kernels for manifold-valued data.
• To enable the application of established machine learning algorithms to Riemannian manifolds.

Main Methods:

• Defined Gaussian radial basis function (RBF)-based positive definite kernels on manifolds.
• Developed a unified framework for analyzing the positive definiteness of Gaussian RBF kernels.
• Applied kernels to the manifolds of symmetric positive definite matrices and Grassmann manifolds.

Main Results:

• Successfully embedded manifold data into a reproducing kernel Hilbert space.
• Identified positive definite kernels for specific Riemannian manifolds.
• Demonstrated the generalization of Euclidean algorithms to manifold data.

Conclusions:

• The proposed positive definite Gaussian kernels effectively handle manifold-valued data.
• This approach allows the use of standard machine learning tools on complex geometric data.
• Enables advanced analysis for computer vision tasks involving Riemannian manifolds.