[Paper Review] QMagFace: Simple and Accurate Quality-Aware Face Recognition

Philipp Terhörst, Malte Ihlefeld+5 more|arXiv (Cornell University)|Nov 26, 2021

Face recognition and analysis4 citations

TL;DR

QMagFace proposes a simple yet effective face recognition method that enhances accuracy by integrating model-specific face image quality into the comparison score via a linear quality-weighting function. Trained with a magnitude-aware angular margin loss (MagFace), it achieves state-of-the-art performance on benchmarks like AgeDB (98.50%), XQLFQ (83.95%), and CFP-FP (98.74%), especially under challenging conditions such as cross-pose and cross-age.

ABSTRACT

Face recognition systems have to deal with large variabilities (such as different poses, illuminations, and expressions) that might lead to incorrect matching decisions. These variabilities can be measured in terms of face image quality which is defined over the utility of a sample for recognition. Previous works on face recognition either do not employ this valuable information or make use of non-inherently fit quality estimates. In this work, we propose a simple and effective face recognition solution (QMagFace) that combines a quality-aware comparison score with a recognition model based on a magnitude-aware angular margin loss. The proposed approach includes model-specific face image qualities in the comparison process to enhance the recognition performance under unconstrained circumstances. Exploiting the linearity between the qualities and their comparison scores induced by the utilized loss, our quality-aware comparison function is simple and highly generalizable. The experiments conducted on several face recognition databases and benchmarks demonstrate that the introduced quality-awareness leads to consistent improvements in the recognition performance. Moreover, the proposed QMagFace approach performs especially well under challenging circumstances, such as cross-pose, cross-age, or cross-quality. Consequently, it leads to state-of-the-art performances on several face recognition benchmarks, such as 98.50% on AgeDB, 83.95% on XQLFQ, and 98.74% on CFP-FP. The code for QMagFace is publicly available

Motivation & Objective

• Address the challenge of inconsistent face recognition performance under unconstrained conditions such as varying illumination, pose, and expression.
• Overcome limitations of prior methods that either ignore face image quality (FIQ) or use non-ideal quality estimates during comparison.
• Develop a generalizable, quality-aware comparison function that leverages model-specific FIQ to improve matching accuracy.
• Demonstrate consistent performance gains across diverse benchmarks and under challenging scenarios like cross-pose, cross-age, and cross-quality.
• Ensure robustness and generalizability across different backbone architectures, training databases, and evaluation protocols.

Proposed method

• Train a face recognition model using the magnitude-aware angular margin (MagFace) loss, which produces embedding magnitudes that correlate with face image quality.
• Extract model-specific face image quality scores from the magnitude of the learned embeddings, enabling quality-aware comparison.
• Introduce a linear quality-weighting function $ \omega(s) $ that modulates the comparison score $ s $ based on the average quality of the two faces.
• Define the quality-aware comparison score as $ \hat{s}(s, q_1, q_2) = s + \omega(s) \cdot \frac{q_1 + q_2}{2} $, where $ \omega(s) $ is learned from optimal thresholds on validation data.
• Optimize the quality-weighting function $ \omega(s) $ using a linear fit to the optimal weights $ \omega_{\text{opt}} $ at various false match rate (FMR) thresholds.
• Apply the quality-aware score during inference to improve separability between genuine and imposter distributions, especially for low-quality or ambiguous face pairs.

Experimental results

Research questions

• RQ1 Can integrating model-specific face image quality into the comparison process consistently improve face recognition performance under unconstrained conditions?
• RQ2 Is a linear function sufficient to model the optimal quality weighting for comparison scores across different databases and decision thresholds?
• RQ3 Does the proposed quality-aware scoring mechanism generalize across different backbone architectures and training datasets?
• RQ4 How does the quality-aware approach perform under challenging recognition scenarios such as cross-pose, cross-age, or cross-quality?
• RQ5 Can the quality-aware comparison function achieve state-of-the-art performance on standard face recognition benchmarks?

Key findings

• QMagFace achieves a recognition accuracy of 98.50% on AgeDB, significantly outperforming prior methods under cross-age and cross-pose conditions.
• On the XQLFQ benchmark, QMagFace reaches 83.95% accuracy, demonstrating strong performance in cross-quality recognition tasks.
• On the CFP-FP benchmark, QMagFace attains 98.74% accuracy, marking a state-of-the-art result for image-based face recognition.
• The quality-aware comparison function consistently improves performance across a wide range of false match rates, especially in the low-FMR regime.
• The optimal quality-weighting function $ \omega(s) $ is linearly correlated with decision thresholds across all tested databases and MagFace models, validating the simplicity and generalizability of the approach.
• The method shows high generalizability: the quality-weighting function learned on one database generalizes well to others, indicating robustness across training and evaluation setups.