1. MULTI-TASK POSE-INVARIANT FACE RECOGNITION
ABSTRACT
Face images captured in unconstrainedenvironments usually contain significant pose
variation,which dramatically degrades the performance of algorithmsdesigned to recognize
frontal faces. This paper proposes a novelface identification framework capable of handling the
full rangeof pose variations within ±90° of yaw. The proposed frameworkfirst transforms the
original pose-invariant face recognitionproblem into a partial frontal face recognition problem. A
robustpatch-based face representation scheme is then developed torepresent the synthesized
partial frontal faces. For each patch,a transformation dictionary is learnt under the proposed
multitasklearning scheme. The transformation dictionary transformsthe features of different
poses into a discriminative subspace.Finally, face matching is performed at patch level rather
thanat the holistic level. Extensive and systematic experimentationon FERET, CMU-PIE, and
Multi-PIE databases shows thatthe proposed method consistently outperforms single-task-based
baselines as well as state-of-the-art methods for the poseproblem. We further extend the
proposed algorithm for theunconstrained face verification problem and achieve top-
levelperformance on the challenging LFW data set.