[Paper Review] MagFace: A Universal Representation for Face Recognition and Quality Assessment

Qiang Meng, Shichao Zhao+2 more|arXiv (Cornell University)|Mar 11, 2021

Face recognition and analysis45 references37 citations

TL;DR

MagFace learns a universal face embedding where the feature magnitude serves as a quality indicator, guiding adaptive margins to improve recognition, quality assessment, and clustering without extra quality labels.

ABSTRACT

The performance of face recognition system degrades when the variability of the acquired faces increases. Prior work alleviates this issue by either monitoring the face quality in pre-processing or predicting the data uncertainty along with the face feature. This paper proposes MagFace, a category of losses that learn a universal feature embedding whose magnitude can measure the quality of the given face. Under the new loss, it can be proven that the magnitude of the feature embedding monotonically increases if the subject is more likely to be recognized. In addition, MagFace introduces an adaptive mechanism to learn a wellstructured within-class feature distributions by pulling easy samples to class centers while pushing hard samples away. This prevents models from overfitting on noisy low-quality samples and improves face recognition in the wild. Extensive experiments conducted on face recognition, quality assessments as well as clustering demonstrate its superiority over state-of-the-arts. The code is available at https://github.com/IrvingMeng/MagFace.

Motivation & Objective

• Motivate end-to-end face recognition systems to account for input face quality without separate quality labels.
• Introduce magnitude as a quality indicator within the face embedding space.
• Develop MagFace loss with magnitude-aware angular margins and a regularizer to structure within-class distributions.
• Demonstrate improvements in face recognition, quality assessment, and clustering on standard benchmarks.

Proposed method

• Define magnitude a_i = ||f_i|| and two auxiliary functions: magnitude-aware margin m(a_i) and regularizer g(a_i).
• Extend ArcFace by replacing fixed margin with m(a_i) and adding regularizer term lambda_g g(a_i).
• Prove two properties: convergence (L_i is strictly convex with a unique optimum a_i^*) and monotonicity (a_i^* increases as cos-distance to the center decreases).
• Argue that larger magnitudes correspond to higher face quality and better recognition likelihood, while low-quality samples are pulled toward the origin within a constrained feasible region.

Experimental results

Research questions

• RQ1 Can feature magnitude be used as a reliable, label-free indicator of face quality within a recognition model?
• RQ2 Does a magnitude-aware loss improve intra-class structure and inter-class separation under unconstrained conditions?
• RQ3 Can MagFace improve downstream tasks such as quality assessment and clustering in addition to recognition?
• RQ4 How do magnitude-aware margins adapt to sample difficulty and prevent overfitting to noise?

Key findings

• MagFace achieves top verification accuracy on multiple benchmarks (LFW, CFP-FP, AgeDB-30, CALFW, CPLFW, IJB-B, IJB-C) compared to ArcFace and other baselines.
• The magnitude of learned features correlates with recognition difficulty and class-center proximity, enabling quality-aware weighting without extra labels.
• MagFace yields improved clustering performance across K-means, AHC, DBSCAN, and L-GCN, outperforming ArcFace baselines on F-score and NMI metrics.
• Qualitative analyses show mean faces become more detailed with higher magnitudes, illustrating MagFace’s quality representation.
• MagFace+ (magnitude-weighted aggregation) further improves identity embeddings on IJB-B/C benchmarks.
• MagFace’s magnitude-based quality assessment outperforms several image- and face-based baselines across multiple datasets.