| Popis: |
Face recognition has attracted wide attention in recent years. The major issue in face recognition is to improve the robustness under various modalities, such as different illuminations, poses, or backgrounds. Due to the rapid development of deep learning, tremendous progress has been achieved. The mainstream to address the problem in face recognition is based on an intuitive objective: maximizing the inner-class similarity as well as the intra-class discrepancy. In this paper, inspired by the philosophy behind contrastive learning, we modify the InfoNCE loss which is usually used in unsupervised representation learning to achieve the objective while theoretically, a lower bound of the mutual information between the input faces of a certain identity and their corresponding representation is maximized. Besides, via our modification, there is a natural alignment of the representations under different modalities of a certain identity. Moreover, the experiment results based on the benchmark face recognition dataset empirically demonstrate the superiority of our method. |
