Design of loss function that enhance discriminative power.
ArcFace: Additive Angular Margin Loss, has a clear geometric interpretation due to the geodesic distance on the hypersphere.
DCNNs map face image into a feature that has small intra-class and large inter-class distance.
Two lines to train:
- multi-class classifier, separates different identities using softmax
- learn an embedding, triplet loss(anchor, positive, negative)
==> Both have drawbacks
W∈Rd×n
- Softmax loss: increases linearly with n; learned feature separable for closed-set classification but not discriminative
- Triplet loss: combinatorial explosion; semi-hard sample mining difficult
Enhanced softmax:
- Centre loss by Wen et al., obtain intra-class compactness; updating centres is difficult.
- multiplicative angular margin penalty, enforce intra-class compactness and inter-class discrepancy, leading to better discriminative. Angular margin, Sphereface.
- CosFace, adds cosine margin penalty to the target logit
This paper: ArcFace, further improve discrimanative power of face recognition & stabilise training process. Dot product between feature & last fc layer equal to the cosine distance after feature & weight norm. Use arc-cosine to cal the angle between current feature & target weight. Add an additive angular margin to target angle, get target logit back by cosine. Re-scale logits by fixed feature norm, subsequent same as softmax.
Advantages:
- Engaging, directly optimises the geodesic distance
- Effective, achieves sota perfomance
- Easy, easy to implement
- Efficient, adds negligible computational complexity
Softmax loss:
L1=−N1i=1∑Nlog∑j=1neWjTxi+bjeWy(i)Txi+by(i)
- Does not explicitly optimise the feature embedding, result in performance gap under large intra-class variations and large-scale test scenarios
SV-Softmax, inherit the advantages of mining-based and margin-based losses into one framework.
Mainly two new methods:
- CVC training strategy that enhances IvS performance
- DP-Softmax(dominant prototype) to make scalable on large-scale classes
Main Formulas:
x=xage⋅xidxage=∥x∥2xid={∥x∥2x1,∥x∥2x2,⋯,∥x∥2xn}Main purpose: Look the L2 length of represetation as age
By Liu Yu
Main Aim: Get the quality of image as weight to build a set represetation
by Wang Zhongdao
GCN method to cluster face images
CASIA WebFace
Collect From IMDB,
IJB Dataset