(Part 1), we discussed about Cosface.
In this part we’ll discuss about Arcface: Additive Angular Margin Loss for Deep Face Recognition.
Introduction
There are two main lines research to train CNN for face recognition, one that train a multi-class classifier using softmax classifier and the other that learn the embeddings such as the triplet loss. However both have their drawbacks:
For the softmax loss, the more you add the different identities for recognition, the more number of parameters will increase. And for the triplet loss there is a combinatorial explosion in the number of face triplets for large scale dataset leading to large number of iterations.
An additive angular margin loss is proposed in arcface to further improve the descriminative power of the face recognition model and stabilize the training process. The arc-cosine function is used to calculate angle between the current feature and target weight. ArcFace directly optimizes the geodesic distance margin by virtue of exact correspondence between angle and arc in the normalized hypersphere.

#computer-vision #deep-learning #face-recognition #loss-function #artificial-intelligence