Motivation
- 通过深度网络学习人脸embedding-feature![DeepID-2014CVPR];
- 减小内类variations,增加类间difference;
Novelty
- identification loss
- verification loss [cite]
, ,d = cos(f1, f2)
- constrastive
Loss = Ident(f, l) + lambda * Verif(f1, f2, l1, l2)
Details
Experiment
- 量化分析
- 方差随着dim的增加衰减得越快,意味着类内的diversity、类间的variation越小【即:pattern比较固定】.
- 当lambda=0时,采用softmax-loss导致类内diversity较大、类间variation较小,因此造成识别干扰!
- 当lambda=0.05时,类内的diversity明显减小、类间variation明显增大,意味学到人脸特征区分能力得到提升!
- 当lambda=∞时,类内的diversity、类间variation均显著增大,因为没有 identification的反馈信息,模型无法获取具有足够多样化信息的嵌入feature,导致模型识别能力锐减!
- 实验结果
Reference
[1]. Deep Learning Face Representation by Joint Identification-Verification [2014-NIPS]
[2]. Deep Learning Face Representation from Predicting 10,000 Classes [2014-CVPR]