Learning Steadily: Accumulating Relative Point Margin Scores for Face Image Quality Assessment

Guray Ozgur, Jan Niklas Kolf, Eduarda Caldeira, Tahar Chettaoui, Fadi Boutros, Naser Damer

Keywords: Face Image Quality, Face Recognition

Abstract: Face Image Quality Assessment (FIQA) is essential for reliable biometric systems, as it determines whether captured face images are suitable for automated recognition tasks. Current state-of-the-art FIQA approaches that integrate with face recognition (FR) training often suffer from unstable quality targets due to the evolving feature space during training. We introduce CARPM-FIQA, a novel approach that addresses this limitation by learning to predict accumulated relative point margin scores across training epochs rather than relying on single-epoch estimates. Our method quantifies image quality by tracking and accumulating the ratio between intra-class compactness and inter-class separation for each sample throughout training, providing a temporally-stable measure of sample utility. We demonstrate theoretically and empirically that this cumulative averaging approach significantly reduces variance in quality estimates and improves reliability in sample ranking compared to non-cumulative alternatives. To enable quality assessment for unseen images, we extend standard FR architectures with a regression layer that predicts these accumulated values. Through extensive evaluation on eight challenging benchmarks (LFW, AgeDB-30, CFP-FP, CALFW, Adience, CPLFW, XQLFW, and IJB-C) and across four state-of-the-art FR models, we show that CARPM-FIQA consistently ranks among the top-performing methods, with particular strength on datasets with quality variations. Our approach offers a more robust alternative to existing FIQA techniques while maintaining tight integration with the FR pipeline.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5784