ArcFace: Additive Angular Margin Loss for Deep Face Recognition
Introduction
- Centre loss penalises the distance between the deep features and their corresponding class centers in Euclidean space to achieve intra-class compactness.
- SphereFace assumes the linear transformation matrix in the last fully connected layer can be used as a representation of the class centers in an angular space and pensalises the angles between the deep features and their corresponding weights in a multiplicative way.
- ArcFace is an additive angular margin loss.
Architecture
Softmax Loss
- Softmax loss doesn’t enforce higher similarity for intra-class samples and diversity for inter-class samples. (See figure below)
ArcFace Loss
-
We change W.x as W x cos\theta and set bias as 0. -
If we normalize W as a unit vector, then W becomes 1. - We also normalize x and scale it to s.
Results
Kaushik Rangadurai
Code. Learn. Explore