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

Learning Steadily. Accumulating Relative Point Margin Scores for Face Image Quality Assessment. By Guray Ozgur and six others, from the Fraunhofer Institute for Computer Graphics Research and the Technical University of Darmstadt. Published on the fifteenth of September. No code repository is linked to it.

Face image quality assessment decides whether a captured face image is suitable for automated face recognition, and reliable biometric systems depend on that judgment. The strongest current methods integrate quality assessment with face recognition, and they share a weakness the authors call temporal instability. While the network trains, its feature space keeps evolving. A quality estimate taken in any single epoch therefore fluctuates from one epoch to the next. The quality target the model is meant to learn keeps moving, and a moving target undermines reliable quality prediction. The problem is not that any one estimate is wrong, but that none of them stays put long enough to be trusted.

CARPM-FIQA is a stabilization strategy for this integrated setting. The quantity it works with is the relative point margin: the ratio between intra-class compactness and inter-class separation. Instead of relying on that measurement from one epoch, it accumulates it across the entire training trajectory. A single snapshot is replaced by a cumulative average over the whole of training.

That cumulative averaging is presented as theoretically grounded, with three advantages claimed for it. Quality estimates have reduced variance. Mean squared error improves. And ranking stability is enhanced, with convergence guarantees as training progresses. The intuition is the familiar one for averaging: individual epochs may wobble as the features shift, but the accumulated value settles. Each new epoch still contributes its measurement, but it no longer gets to overwrite everything that came before it, and the estimate the model learns from becomes steadier the longer training runs. Nothing about the margin itself is redefined. What changes is how long a memory the quality target has. Two variants are evaluated, labelled L and S.

Controlled experiments use the SynFIQA dataset, whose images carry labelled quality groups. There, cumulative averaging shows superior discriminative ability, and ablation studies across different training configurations confirm consistent improvements. That is the direct test of the idea: the same quality measure, once averaged over training and once taken from a single epoch, compared on data where the true quality groups are known.

The broad comparison is large. CARPM-FIQA is evaluated against twelve FIQA methods, on eight challenging benchmarks, with four face recognition models, at two FMR thresholds. Ranked by pAUC-EDC and AUC-EDC, averaged across recognition models and then, after normalizing each benchmark, across benchmarks, the L variant places fourth and the S variant sixth, out of seventeen compared methods. Both stay within a few percent of the best method's normalized average for every recognition model.

The claim here is not top rank. It is stability, gained without giving up the performance benefits that integration with recognition brings: a principled answer to training instability that keeps results close to the best. The authors also draw a wider lesson. Wherever a training target fluctuates because the feature representations underneath it are still evolving, aggregating over time can steady the objective.