Differential Gaze Estimation with Ocular Counter-Rolling Compensation

The main challenges in the gaze estimation task are the variations in head poses, eye shapes and inner eye structures amongst individuals. Conventional methods have proposed many techniques to reduce these variations. One of the most adopted techniques is the normalization method which realigns the head pose to remove the roll component of head pose. However, this global transformation, which forces the eye to move with the other part of the face together, violates the human anatomy as eyeballs are controlled separately by the surrounding muscles. These muscles restrain the eyeball to be horizontally stable even when the head is tilted in a natural pose. In this work, we propose a novel method to model this counter-rolling effect and use it to remove the error from the normalization step. Experiments validate our approach, which outperforms state-of-the-art eye-image-only gaze estimation methods.