# ArcFace: Additive Angular Margin Loss for Deep Face Recognition

**Authors:** Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, Stefanos Zafeiriou  
**arXiv:** [1801.07698](https://arxiv.org/abs/1801.07698)  
**Date:** 2018-01-23  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.07698v4/

## Key points

- Incorporating margins into softmax loss function
- Clear geometric interpretation
- Boosts discriminative power
- Encourages one dominant sub-class with clean faces
- Non-dominant sub-classes for hard or noisy faces
- Pre-trained ArcFace model can generate identity-preserved face images using network gradient and BN priors
- LFR2019-Video and LFR2019-Image dataset details
- Effectiveness of ArcFace in enhancing face recognition tasks through margin optimization and feature synthesis.

## Summary

In the field of face recognition, a popular research direction involves incorporating margins into the softmax loss function to improve class separability. This paper introduces the Additive Angular Margin Loss (ArcFace), which offers a clear geometric interpretation and significantly boosts discriminative power. To address label noise issues, the concept of sub-center ArcFace is proposed. Each class contains multiple sub-centers for training samples to be close to, encouraging one dominant sub-class with clean faces and non-dominant sub-classes containing hard or noisy faces. This leads to enhanced performance in purifying raw web faces under real-world noise. Furthermore, the study explores mapping feature vectors to face images without requiring additional generator or discriminator training. The pre-trained ArcFace model can generate identity-preserved face images for subjects both within and outside the training data using network gradient and Batch Normalization (BN) priors. Extensive experiments demonstrate that ArcFace not only improves discriminative feature embedding but also strengthens generative face synthesis capabilities. Moreover, detailed dataset statistics are provided for large-scale video datasets such as LFR2019-Video and the LFR2019-Image dataset comprising 274K images from 5.7K LFW identities and 1.58M distractors downloaded from Flickr. The study showcases the effectiveness of ArcFace in enhancing face recognition tasks through its innovative approaches to margin optimization and feature synthesis. techniques have been widely studied in recent years with a focus on improving  methods in order to enhance . One such method is the Additive Angular Margin Loss (ArcFace), which not only has a clear geometric interpretation but also significantly boosts performance. To tackle label noise issues, this study proposes a new approach called sub-center ArcFace, which encourages one dominant sub-class with clean faces and non-dominant sub-classes containing hard or noisy faces. This leads to better performance in purifying raw web faces under real-world noise. Additionally, the study explores a new method for mapping feature vectors to face images without requiring additional training of generators or discriminators. The pre-trained ArcFace model can generate identity-preserved face images for subjects both within and outside the training data using network gradient and Batch Normalization (BN) priors. Extensive experiments demonstrate that ArcFace not only improves discriminative feature embedding but also strengthens generative face synthesis capabilities. Furthermore, detailed dataset statistics are provided for large-scale video datasets such as LFR2019-Video and the LFR2019-Image dataset comprising 274K images from 5.7K LFW identities and 1.58M distractors downloaded from Flickr. These datasets showcase the effectiveness of ArcFace in enhancing face recognition tasks through its innovative approaches to margin optimization and feature synthesis.