Using overlapping distributions to deal with face pose mismatch

Abstract

Representing the face as a distribution of freely moving patches, which we refer to as a "free-parts" representation, has recently demonstrated some benefit in the task of face verification. This benefit can be largely attributed to the representation's natural ability to deal with local appearance variation within the face. Hitherto, a major limitation that has hindered the wider adoption of this type of facial representation, for the task of face verification, has been its poor ability to take advantage of prior knowledge concerning mismatches in context; such as pose. This paper goes some way to alleviating these limitations by making two novel contributions: (i) Demonstrating that free-parts distributions of a client's face for different poses overlap to such a degree that a considerable amount of discrimination is preserved in the intersection. (ii) Through the off-line estimation of subject-independent pose dependent priors, an alternative to the canonical log-likelihood measure can be employed that takes advantage of this intersection and is less sensitive to mismatch in the presence of pose variation.