ArcFace: Additive Angular Margin Loss for Deep Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Softmax loss functions in face recognition are being enhanced with margins to improve class separability.
• Existing methods may be susceptible to label noise in large-scale datasets.

Purpose of the Study:

• Introduce Additive Angular Margin Loss (ArcFace) for enhanced discriminative power in face recognition.
• Propose sub-center ArcFace to mitigate the impact of massive label noise.
• Explore the generative capabilities of ArcFace for face synthesis.

Main Methods:

• Developed Additive Angular Margin Loss (ArcFace) with a clear geometric interpretation.
• Introduced sub-center ArcFace, where each class has K sub-centers to handle noisy or hard samples.
• Utilized network gradients and Batch Normalization priors for identity-preserving face image generation without additional generators or discriminators.

Main Results:

• ArcFace significantly enhances the discriminative power of feature embeddings.
• Sub-center ArcFace effectively purifies raw web faces under massive real-world noise, boosting performance.
• Pre-trained ArcFace models can generate identity-preserved face images.

Conclusions:

• ArcFace provides a robust approach to discriminative feature embedding in face recognition.
• Sub-center ArcFace offers a solution for handling label noise in large-scale face datasets.
• ArcFace demonstrates dual capabilities in both feature embedding and generative face synthesis.