Two Kinds of Statistics for Better Face Recognition

Abstract

We briefly review the base techniques of elastic graph matching [1] and elastic bunch graph matching [2], which provide a method for face detection, matching, comparison, and identity decision. We then present a method that combines the advantages of Gabor-labeled graphs with maximum likelihood decision making. The improvements over pure bunch graph matching have been studied, and the method has been successfully applied to large face databases like the one for the Face Recognition Grand Challenge [3]. Finally, we describe a system based on rank order statistics that can learn invariances in a minimally supervised way from a set of examples of individual faces in several situations like different head pose or illumination. Recognition rates are improved significantly without explicit modeling of the image formation process [4].