Similarity Rank Correlation for Face Recognition Under Unenrolled Pose

Face recognition systems have to deal with the problem that not all variations of all persons can be enrolled. Rather, the variations of most persons must be modeled. Explicit modeling of different poses is awkward and time consuming. Here, we present a subsystem that builds a model of pose variation by keeping a model database of persons in both poses, additionally to the gallery of clients known in only one pose. An identification or verification decision for probe images is made on the basis of the rank order of similarities with the model database. Identification achieves up to 100% recognition rate on 300 pairs of testing images with 45 degrees pose variation within the CAS-PEAL database, the equal error rate for verification reaches 0.5%.