SIGNATURE WORK
CONFERENCE & EXHIBITION 2022

Gaze and gaze zone estimation has wide application in VR/AR and social robotics for healthcare, medical treatment, and education. For example, a social robot can infer the users’ intention through their gazing area; a camera can make conclusions about the products’ popularity through the customers’ attention; car companies can provide warning prompts when the drivers are not focusing. However, the performance of the gaze estimation algorithms is suppressed by their sensitivity to different illuminations, subject identities, and viewing angles, and the performance of the models trained on different datasets is hard to be compared. Therefore, this signature work project studies these two obstacles to the real-world application of gaze estimation algorithms. To tackle the lack of robustness, we propose a variance and invariance learning framework for generalizable gaze estimation, whose effectiveness is evaluated by the models’ angular error on the public dataset ETH-XGaze. To tackle the lack of a universal evaluation metric, we propose a multi-view multi-screen 3D gaze reconstruction system, where three screens, three cameras, and the subject’s gaze are visualized in one world coordinate system. Because of this system and our collected test videos, we can better access an algorithm’s performance quantitively through our presented gaze point error and qualitatively through visualization.