Outlier Constrained Unsupervised Domain Adaptation Algorithm for Gaze Estimation

Abstract

In recent years, gaze estimation has been applied to numerous application areas, such as driver monitor system, autism assessment, and so on. However, current practical gaze estimation algorithms require a large amount of data to obtain better results. The collection of gaze data requires specific equipment, and the collection process is cumbersome, tedious and lengthy. Moreover, in some scenarios, like the autism assessment scenario, it is impossible to obtain the gaze training data of autistic children due to their social communication disorders. Therefore, we need to generalize a model trained on public datasets to a new scenario without gaze ground truth labels. In this study, we tackle this problem by leveraging adversarial learning to implement domain adaptation. Besides, we propose an outlier loss to supervise the outputs of the target domain. We test our domain adaptation algorithm on the XGaze-to-MPII domain adaptation task, and achieve a performance improvement of 14.7%.