Implicit space pose consistent transfer network for deep face verification

Abstract

Unconstrained face recognition has achieved state-of-the-art performance with the help of the powerful representation capabilities of deep convolutional neural networks. However, face recognition always suffers from other intra-class imbalances due to the influence of pose variations present in the real world, which significantly degrades the performance, especially for tilted face recognition under surveillance videos, which is rarely studied. Tilted face recognition is more challenging than frontal and profile face recognition due to the effect of self-obscuration. To this end, we consider promoting the frontalization of tilted faces in the implicit space, while preserving identity information. In this paper, we propose an implicit spatial pose consistent transfer network (PCTN) that exploits the prior guidance of pose information to enhance the model’s ability to perceive pose. Specifically, we first develop the multi-path attention block (MPAB), which is used as the main block for feature embedding in deep networks. Such a design is employed to enhance the relevance of multi-layer convolutional features and provide a powerful foundation for the deep asymptotic transfer of feature frontalization. Then, the pose consistent transfer module (PCTM) combines pose labels and deep embedding features to learn robust discriminative information. PCTM consists of two parts: the pose guidance block (PGB) explores the implicit nonlinear feature mapping from pose faces to frontal faces in feature space, and the attention compensation module (ACM) aims to exploit the complementarity of dense and sparse features, enhancing pose-related features and suppressing pose-unrelated features in terms of both spatial-wise and channel-wise. Furthermore, we investigate the performance comparison on datasets with multiple pose variations without performance degradation. Extensive experimental results on multiple benchmark datasets demonstrate the effectiveness of our proposed method.