A New Distance Criterion for Face Recognition Using Image Sets

A major face recognition paradigm involves recognizing a person from a set of images instead of from a single image. Often, the image sets are acquired from a video stream by a camera surveillance system, or a combination of images which can be non-contiguous and unordered. An effective algorithm that tackles this problem involves fitting low-dimensional linear subspaces across the image sets and using a linear subspace as an approximation for the particular face identity. Unavoidably, the individual frames in the image set will be corrupted by noise and there is a degree of uncertainty on how accurate the resultant subspace approximates the set. Furthermore, when we compare two linear subspaces, how much of the distance between them is due to inter-personal differences and how much is due to intrapersonal variations contributed by noise? Here, we propose a new distance criterion, developed based on a matrix perturbation theorem, for comparing two image sets that takes into account the uncertainty of estimating a linear subspace from noise affected image sets.