该文采用将特征或者权重矩阵归一化的方法,并引入了agent vecotor 代替了hard mining。论文中将方法在别人模型上进行finetuning,比自己从头训模型的效果要更好。
用来finetuning的模型:
Wu’s model is a 10-layer plain CNN with Maxout[6] activation unit.
https://github.com/AlfredXiangWu/face_verification_experiment
Wen’s model is a 28-layer ResNet[7] trained with both softmax loss and center loss.