Abstract:
Face recognition has achieved remarkable improvements with the help of the angular margin based softmax losses. However, the margin is usually manually set and kept constant during the training process, which neglects both the optimization difficulty and the informative similarity structures among different instances. Although some works have been proposed to tackle this issue, they adopt similar methods by simply changing the margin for different classes, leading to limited performance improvements. In this paper, we propose a novel sample-wise adaptive margin loss function from the perspective of the hypersphere manifold structure, which we call companion guided soft margin (CGSM). CGSM introduces the information of distribution in the feature space, and conducts teacher-student optimization within each mini-batch. Samples of better convergence are considered as teachers, while students are optimized with extra soft penalties, so that the intra-class distances of inferior samples can be further compacted. Moreover, CGSM does not require sophisticated mining techniques, which makes it easy to implement. Extensive experiments and analysis on MegaFace, LFW, CALFW, IJB-B and IJB-C demonstrate that our approach outperforms state-of-the-art methods using the same network architecture and training dataset.
