Gaze Data Imbalance: An Overlooked Challenge in Appearance-Based Gaze Estimation

Data imbalance exists in appearance-based gaze estimation datasets, hurting the model’s generalizability and fairness. Unfortunately, this aspect is usually overlooked, and researchers focus mainly on developing more and more sophisticated gaze estimation neural networks. In this work, we identify two types of imbalance in gaze estimation data. The first is related to the uneven distribution of ground truth gaze vectors, and the second one comes from the uneven ethnicity distribution of dataset participants. We prove the negative impact of both of them on the model’s generalizability. Therefore, we propose Uniform Gaze Sampling and Uniform Ethnicity Sampling, simple yet effective re-sampling techniques tailored for gaze estimation. Moreover, we introduce balanced metrics, i.e., Balanced Gaze Error and Balanced Ethnicity Error, for a fair performance evaluation. Finally, we demonstrate the usefulness of the proposed methods and metrics on four benchmarks. To the best of our knowledge, we are the first to address data imbalance in gaze estimation comprehensively.