APA: Adaptive Pose Alignment for Pose-Invariant Face Recognition

Face alignment is a crucial step in face recognition. Through making the position of face consistent, face alignment reduces intra-class variability due to factors such as lighting, background, pose, and perspective transformation, and further facilitating the recognition tasks. In this paper, we propose a new face alignment method for pose-invariant face recognition, called adaptive pose alignment (APA), which can greatly reduce the intra-class difference and correct the noise caused by the traditional method in the alignment process, especially in unconstrained settings. Instead of aligning all faces to the pre-defined, uniform frontal shape, we adaptively learn the alignment templates according to the facial poses and then align each face of training or testing sets to its related template. To further improve face recognition performance, we propose a simple, yet effective feature normalization method which can generate more discriminative feature representation of a face or a set of faces (template) combined with the APA method. Furthermore, we introduce a pose-invariant face recognition pipeline that sequentially applies APA-based alignment, deep representation by softmax or ArcFace loss function, and the effective feature normalization procedure. We empirically show that APA-based images can accelerate the training of deep face recognition model by aligning all the images to the optimal templates. Moreover, experimental results show that the proposed method achieves the state-of-the-art performance on popular challenging face datasets including IJB-A, IJB-C, and CPLFW datasets.