Sphere Margins Softmax for Face Recognition

In conjunction with the cross-entropy loss, softmax function has gained great popularity in supervised deep convolutional neural networks (DCNNs). Although original softmax loss reduces training difficulty and makes multi classification problems easier to converge, the module does not explicitly encourage compactness within class and separability between classes, from this point of view it is not particularly suitable for face recognition tasks. In this paper, we reformulate the softmax loss with sphere margins (SM-Softmax) by normalizing both weights and extracted features of the last fully connected layer and have quantitatively adjustable angular margin by hyperparameter m
1
and m
2
. Extensive experiments on CASIA-WebFace and Labeled Face in the Wild (LFW) validate that our SM-Softmax gives better results than the present state-of-the-art methods while adopting the same experimental configuration (benchmark datasets and network structure) and the convergence speed in the early stage is accelerated significantly.