cosface: large margin cosine loss for deep face recognition

Wang H, Wang Y, Zhou Z, et al. CosFace: Large Margin Cosine Loss for Deep Face Recognition[C]. computer vision and pattern recognition, 2018: 5265-5274.

@article{wang2018cosface:,
title={CosFace: Large Margin Cosine Loss for Deep Face Recognition},
author={Wang, Hao and Wang, Yitong and Zhou, Zheng and Ji, Xing and Gong, Dihong and Zhou, Jingchao and Li, Zhifeng and Liu, Wei},
pages={5265–5274},
year={2018}}

概

本文从angular margin角度提出了对交叉熵损失的一个改进.

主要内容

一般的softmax交叉熵损失为
$$L_s=\frac{1}{N}\sum _{i=1}^N-\log \frac{e^{f_{y_i}}}{{\sum }_{j=1}^C{e}^{f_{y_j}}},$$
其中
$$f_j=W_j^{T}x=\Vert W_j\Vert \Vert x\Vert \cos {\theta }_j,$$

固定$\Vert W_j\Vert =1,\Vert x\Vert =s$, 则
$$L_{ns}=\frac{1}{N}\sum _i-\log \frac{e^{s\cos ({\theta }_{y_i},i)}}{{\sum }_j{e}^{s\cos ({\theta }_{y_j},i)}}$$
只与角度angular margin有关, 所以实际上, 一个类别属于$i$就是当
$$\cos {\theta }_i>\cos {\theta }_j,\forall j\ne i,$$
为了给其增加一些难度, 我们可以
$$\cos {\theta }_i-m>\cos {\theta }_j,\forall j\ne i,$$
即我们在$\cos {\theta }_i>\cos {\theta }_j$的基础上, 进一步要求其angular margin进一步提高, 这就是large angular margin的思想.

于是本文的损失为: