Scalable softmax loss for face verification

Abstract

Thanks to the recent development of deep convolutional neural networks, the performance of face verification has increased significantly. The softmax loss function is used mostly to make CNN models trained well. In order to get more robust face feature by deep CNN model, this paper proposes a new supervision signal based on the regular softmax loss function, namely scalable softmax loss, for face verification task. The scalable softmax loss function adjust the contribution to final loss for different training samples by a learned parameter. And, it's important to note that our proposed scalable softmax loss function can be easily implemented using existing deep learning frameworks. Extensive analysis and experiments on Labeled Face in the Wild(LFW) and YouTube Faces(YTF) show the superiority of the scalable softmax loss function in face verification task. Specially, our proposed scalable achieves comparable results on challenging LFW data set and YTF data set with the accuracy 99% and 95.08% respectively. Codes are released at1.