ArcFace

ArcFace is a loss function used in deep face recognition to enhance the discriminative power of learned embeddings by enforcing an additive angular margin between classes. It operates on feature vectors that are normalized to lie on a hypersphere, as do the class weight vectors in the final fully connected layer, which promotes angular separation rather than purely Euclidean distance.

The core idea is to modify the logit corresponding to the ground-truth class by introducing an angular margin. Given a input feature x and a weight vector W_y for the true class y, the model computes the angle theta between x and W_y. Instead of using the raw cosine similarity cos(theta) for the correct class, ArcFace uses cos(theta + m), where m is a predefined margin. All other classes retain their cos(theta_j) logits. The resulting logits are typically scaled by a factor s before applying softmax cross-entropy loss. This creates a larger angular gap between classes, encouraging more compact intra-class variation and larger inter-class separation.

ArcFace is a successor to earlier margin-based approaches such as SphereFace and CosFace, and is designed to provide stable training and improved accuracy by focusing on angular margins. In practice, the method requires selecting appropriate hyperparameters such as the margin m and the scale s.

History and impact: ArcFace was introduced in 2019 as Additive Angular Margin for Deep Face Recognition and has since become widely adopted in face recognition research and applications, contributing to higher performance on standard benchmarks and robust feature representations for verification and identification tasks.

Limitations and considerations: Like other margin-based losses, ArcFace can require careful tuning for different datasets and backbone architectures, and large-scale training may demand significant computational resources. Ethical considerations in face recognition remain important in deployment.