Fairness meets cross-domain learning: A benchmark of models and metrics

Deep learning-based recognition systems are deployed at scale for real-world applications that inevitably involve our social life. Although of great support when making complex decisions, they might capture spurious data correlations and leverage sensitive attributes (e.g., age, gender, ethnicity). How to factor out this information while maintaining high performance is a problem with several open questions, many of which are shared with those of the domain adaptation and generalization literature which aims at avoiding visual domain biases. In this work, we propose an in-depth study of the relationship between cross-domain learning (CD) and model fairness, by experimentally evaluating 14 CD approaches together with 3 state-of-the-art fairness algorithms on 5 datasets of faces and medical images spanning several demographic groups. We consider attribute classification and landmark detection tasks: the latter is introduced here for the first time in the fairness literature, showing how keypoint localization may be affected by sensitive attribute biases. To assess the analyzed methods, we adopt widely used evaluation metrics while also presenting their limits with a detailed review. Moreover, we propose a new Harmonic Fairness (HF) score that can ease unfairness mitigation model comparisons. Overall, our work shows how CD approaches can outperform state-of-the-art fairness algorithms and defines a framework with dataset and metrics as well as a code suite to pave the way for a more systematic analysis of fairness problems in computer vision (Code available at: https://github.com/iurada/fairness_crossdomain).