A Generic Framework for Evaluating Gaze Representations for Gaze Estimation

While gaze representations have shown impressive performance in gaze estimation tasks, they face significant limitations, as most existing evaluations fail to consider the generalizability of these representations in broader settings. To address this gap, we propose a generic evaluation framework specifically designed to assess the generalizability of gaze representations within the context of gaze estimation. Our framework includes detailed spatial and temporal analyses. For the former, we introduce pairwise distance metrics and linearity assessments to measure the global structure of representations. For the latter, we incorporate metrics for fixation and saccade detection, along with Cohen's Kappa to evaluate agreement. Finally, we present a unified metric that integrates the perspectives mentioned earlier. We perform extensive comparisons across six publicly available baselines, including three supervised and three unsupervised or few-shot learning approaches. Experiments on three benchmarks show that our proposed metrics strongly correlate with representation generalizability, offering an interpretable and effective evaluation tool. When combined with the mutually informative conventional accuracy-based metrics, our approach enables more informed, long-term decision-making for tasks like efficient few-shot calibration or adaptation to unseen domains.