Picture-Specific Cohort Score Normalization for Face Pair Matching

Face pair matching is the task of deciding whether or not two face images belong to the same person. This has been a very active and challenging topic recently due to the presence of various sources of variation in facial images, especially under unconstrained environment. We investigate cohort normalization that has been widely used in biomet-ric verification as means to improve the robustness of face recognition under challenging environments to the face pair matching problem. Specifically, given a pair of images and an additional fixed cohort set (identities of cohort samples never appear in the test stage), two picture-specific cohort score lists are computed and the correspondent score profiles of which are modeled by polynomial regression. The extracted regression coefficients are subsequently classified using a classifier. We advance the state-of-the-art in cohort normalization by providing a better understanding of the cohort behavior. In particular, we found that the choice of the cohort set had little impact on the generalization performance. Furthermore, the larger the size of the cohort set, the more stable the system performance becomes. Experiments performed on the Labeled Faces in the Wild (LFW) benchmark show that our system achieves performance that is comparable to state-of-the-art methods.