Learning Person-Specific Face Representations

Abstract

Humans are natural face recognition experts, far outperforming current automated face recognition algorithms, especially in naturalistic, ``in-the-wild'' settings. However, a striking feature of human face recognition is that we are dramatically better at recognizing highly familiar faces, presumably because we can leverage large amounts of past experience with the appearance of an individual to aid future recognition. Researchers in psychology have even suggested that face representations might be partially tailored or optimized for familiar faces. Meanwhile, the analogous situation in automated face recognition, where a large number of training examples of an individual are available, has been largely underexplored, in spite of the increasing relevance of this setting in the age of social media. Inspired by these observations, we propose to explicitly learn enhanced face representations on a per-individual basis, and we present a collection of methods enabling this approach and progressively justifying our claim. By learning and operating within person-specific representations of faces, we are able to consistently improve performance on both the constrained and the unconstrained face recognition scenarios. In particular, we achieve state-of-the-art performance on the challenging PubFig83 familiar face recognition benchmark. We suggest that such person-specific representations introduce an intermediate form of regularization to the problem, allowing the classifiers to generalize better through the use of fewer --- but more relevant --- face features.