# MagFace: A Universal Representation for Face Recognition and Quality Assessment

**Authors:** Qiang Meng, Shichao Zhao, Zhida Huang, Feng Zhou  
**arXiv:** [2103.06627](https://arxiv.org/abs/2103.06627)  
**Date:** 2021-03-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.06627v4/

## Key points

- MagFace is a novel approach to face recognition and quality assessment
- It addresses the issue of degraded performance when faced with increased variability in acquired faces
- The new loss function ensures that the magnitude of the feature embedding monotonically increases if the subject is more likely to be recognized, making it an effective tool for improving face recognition accuracy
- MagFace introduces an adaptive mechanism that learns well-structured within-class feature distributions by pulling easy samples towards class centers while pushing hard samples away
- Extensive experiments conducted on face recognition, quality assessments, and clustering demonstrate MagFace's superiority over state-of-the-art methods
- The code for implementing MagFace is available at https://github.com/IrvingMeng/MagFace
- The authors of this paper are Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou
- Their work was accepted at CVPR 2021 as an oral presentation.

## Summary

MagFace is a novel approach to face recognition and quality assessment that addresses the issue of degraded performance when faced with increased variability in acquired faces. This problem has been previously tackled by monitoring face quality in pre-processing or predicting data uncertainty along with the face feature. However, MagFace proposes a category of losses that learn a universal feature embedding capable of measuring the quality of a given face. The new loss function ensures that the magnitude of the feature embedding monotonically increases if the subject is more likely to be recognized, making it an effective tool for improving face recognition accuracy. In addition to its innovative loss function, MagFace introduces an adaptive mechanism that learns well-structured within-class feature distributions by pulling easy samples towards class centers while pushing hard samples away. This prevents models from overfitting on noisy low-quality samples and improves face recognition performance in real-world scenarios. Extensive experiments conducted on face recognition, quality assessments, and clustering demonstrate MagFace's superiority over state-of-the-art methods. The code for implementing MagFace is available at https://github.com/IrvingMeng/MagFace. The authors of this paper are Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Their work was accepted at CVPR 2021 as an oral presentation. Overall, MagFace represents a significant advancement in addressing challenges associated with recognizing faces under varying conditions and provides a promising avenue for future research in this area.