Appearance Based Gaze Estimation Using Eye Region Landmarks and Math Approach

Abstract

The gaze direction can be defined by the pupil and the center of the eyeball, the latter cannot be observed in the 2D image, which can cause ill-posed problems and cannot achieve highly accurate gaze estimation. Therefore, we try to extract several effective landmarks around the eyeball and iris from the monocular input for gaze estimation. Instead of directly returning the two angles for the pitch and yaw of the eyeball, we return to an intermediate graphical representation, which in turn simplifies the task of 3D gaze estimation. W e try to use a novel learning-based method to locate the landmarks of an eyeball with the appearance-based method. Through these high-accuracy feature points, we have proposed a new and effective formula for drawing more accurate gaze directions. As for the individual gaze estimation of independent people, our method is superior to existing model fitting and appearance-based methods.