Investigating Bias and Fairness in Appearance-based Gaze Estimation

Significant fairness gaps exist in gaze estimation systems: The paper reveals that state-of-the-art appearance-based gaze estimation models show substantial performance disparities across different ethnic and gender groups, indicating that these systems work better for some demographic groups than others.

This is the first comprehensive fairness benchmark for gaze estimation: The research establishes the first extensive evaluation framework for measuring algorithmic bias in gaze estimation, filling a critical gap since fairness in this technology had been "largely unexplored" despite widespread use in applications like eye-tracking and human-computer interaction.

Standard bias mitigation strategies have limited effectiveness: When the researchers tested existing debiasing techniques on gaze estimation models, they found that "their fairness contributions are limited," suggesting that solutions developed for other AI domains don't automatically solve fairness problems in gaze estimation.

Demographic disparities affect real-world applications: Since gaze estimation is used in accessibility tools, virtual reality, and attention monitoring systems, performance gaps across ethnic and gender groups mean these technologies may work poorly for certain populations, potentially excluding or disadvantaging them.

Open research questions remain: The paper identifies "key insights and open issues," indicating that developing truly equitable gaze estimators requires further research and that the field needs continued attention to fairness as these systems become more widely deployed.