A study on distance measures of tensor manifold for face recognition

Abstract

Gabor-based region covariance matrices (GRCM) or known as tensor are a powerful face image descriptor and have shown promising result in face recognition. The GRCM lie on tensor manifold is inherently non-Euclidean. As such the distance measure on tensor manifold should take the geometry characteristic of the curvature into account. Presently, Affine Invariant Riemannian Metric is the most popular geodesic distance used in literature despite its heavy computation load. This paper studies several alternative distance measures and investigate their tradeoff between performance and computation time.