An Improved Adaptive Angular Margin Loss Function for Deep Face Recognition

In recent years, there have been growing interests on deep learning based face recognition which currently produces state of the art standards in face detection, recognition and verification tasks. As is well known, loss function for extracting face feature plays a crucial role in deep face model. In this regards, margin-based loss functions which apply a fixed margin between the feature and the weight have attracted many interests. However, such margin-based losses have a somewhat limitation in enhancing the discriminative power and generalizability of the face model, since the intra-class and inter-class variations in the real face training sets are often imbalanced. In particular, the embedding feature whose angle between the feature and the weight is distributed around 90° or 180° on the hypersphere reflects the difficult embedding feature in the process of classes. These phenomena occur when one considers those class which contains few number of embedding data. In order to address this problem, in this paper we propose an improved adaptive angular margin loss that incorporates the adaptive and robust angular margin on the angular space between the feature and the corresponding weight instead of constant margin. Our new margin loss function is constructed by incorporating adaptive and more robust angular margin constraint on angular space between the embedding feature and the corresponding weight. The proposed loss function improves the feature discrimination by minimizing the intra-class variation and maximizing the inter-class variation simultaneously. We present some experimental result on LFW, CALFW, CPLFW, AgeDB and MegaFace benchmarks, which demonstrate the effectiveness of the proposed approach.