Abstract
As human eyes serve as conduits of rich information, unveiling emotions,intentions, and even aspects of an individual's health and overall well-being,gaze tracking also enables various human-computer interaction applications, aswell as insights in psychological and medical research. However, existing gazetracking solutions fall short at handling free user movement, and also requirelaborious user effort in system calibration. We introduce Tri-Cam, a practicaldeep learning-based gaze tracking system using three affordable RGB webcams. Itfeatures a split network structure for efficient training, as well asdesignated network designs to handle the separated gaze tracking tasks. Tri-Camis also equipped with an implicit calibration module, which makes use of mouseclick opportunities to reduce calibration overhead on the user's end. Weevaluate Tri-Cam against Tobii, the state-of-the-art commercial eye tracker,achieving comparable accuracy, while supporting a wider free movement area. Inconclusion, Tri-Cam provides a user-friendly, affordable, and robust gazetracking solution that could practically enable various applications.