Addressing Age Bias in the Application of Appearance-Based Gaze-Tracking for Older Adults

Recently, researchers have made great improvements in appearance-based gaze tracking methods using deep learning models. The improvements in accuracy and generalization bring appearance-based gaze tracking technologies towards the ability to be widely used in various applications. Older adults could greatly profit from this easy-to-use and cost-efficient gaze-tracking in areas such as healthcare, assisted devices and gaming applications. However, publicly available datasets for training deep learning models for gaze estimation consist primarily of subjects under the age of 50, creating an age-bias towards younger participants. The results of this study show that older adults had significantly larger fixation errors compared to younger participants when training an appearance-based model with an age-biased dataset. In contrast, when training on a more age-diverse training set, we observed significantly higher improvements in accuracy for older adults compared to Younger participants. To allow a wider audience to take advantage of this promising technology, care must be taken to generate more age-diverse training and testing datasets.