Angle-based softmax loss for face verification

In the past, the design of the loss function for face verification mostly focused on adding different margins, ignoring the gradient of the loss function. Based on the gradient of different loss functions, we propose a loss function that has a better geometric interpretation and uses the angles between the feature vector and the weights of the fully connected layer as a variable to directly optimize the intra-class similarity and the inter-class similarity in this paper. Most importantly, our experiments show that the variable using angle as the loss function can have a more reasonable gradient and improve the accuracy of face recognition.