ArcFace Loss
One of tha main challenges in feature learning using Deep Convolutional Neural Networks for large scale face recognition is designing the optimal loss function that enhance discriminative power.
Before this paper, there were mainly 2 lines to train DCNNs for face recognition.
- Training a multi-class classifier by using a softmax
⇒ Softmax can only provide roughly separable feature embeddings but produces noticeable ambiguity in decision boundaries. -
Learning the direct embeddings, such as the triplet loss which I shared yesterday
⇒ There is a combinatorial explosion in the number of face triplets especially for large-scale datasets, which leads to significant increase in the number of iteration steps.
The ArcFace loss can obviously enforce a more evident gap between the nearest classes.
Reference: ArcFace: Additive Angular Margin Loss for Deep Face Recognition
