Abstract
A major challenge in face recognition is handling large pose variations. Here, we proposed to tackle this challenge by a three step sparse representation based method: estimating the pose of an unseen non-frontal face image, generating its virtual frontal view using learned view-dependent dictionaries, and classifying the generated frontal view. It is assumed that for a specific identity, the representation coefficients based on the view dictionary are invariant to pose and view-dependent frontal view generation transformations are learned based on pair-wise supervised dictionary learning. Experiments conducted on FERET and CMU-PIE face databases depict the efficacy of the proposed method.