Enhancing gaze estimation through fine-grained analysis of eye region

Gaze estimation is the task of estimating the direction of human gaze, and gaze information is an important component of understanding human nonverbal communication. Appearance-based methods have low hardware cost and low hurdles in terms of shooting conditions, since they use only appearance images from common RGB cameras as input. Therefore, appearance-based gaze estimation methods are in high demand in the real world. However, previous appearance-based methods have not adequately taken head orientation into account. Therefore, this paper proposes a new gaze estimation method that includes two block structures, coarse and fine, to account for pupil orientation and head orientation. Evaluation experiments show state-of-the-art performance on the Gaze360, RT-GENE, and MPIIFaceGaze datasets. We confirm that the proposed method, which mimics the human gaze mechanism, is effective in gaze estimation.