MagFace: A Universal Representation for Face Recognition and Quality Assessment

The performance of face recognition system degrades when the variability ofthe acquired faces increases. Prior work alleviates this issue by eithermonitoring the face quality in pre-processing or predicting the datauncertainty along with the face feature. This paper proposes MagFace, acategory of losses that learn a universal feature embedding whose magnitude canmeasure the quality of the given face. Under the new loss, it can be proventhat the magnitude of the feature embedding monotonically increases if thesubject is more likely to be recognized. In addition, MagFace introduces anadaptive mechanism to learn a wellstructured within-class feature distributionsby pulling easy samples to class centers while pushing hard samples away. Thisprevents models from overfitting on noisy low-quality samples and improves facerecognition in the wild. Extensive experiments conducted on face recognition,quality assessments as well as clustering demonstrate its superiority overstate-of-the-arts. The code is available athttps://github.com/IrvingMeng/MagFace.