Block-Wise Gaze Estimation Based on Binocular Images

Appearance-based gaze estimation methods have been proved to be highly effective. Different from the previous methods that estimate gaze direction based on left or right eye image separately, we propose a binocular-image based gaze estimation method. Considering the challenges in estimating the precise gaze points via regression models, we estimate the block-wise gaze position by classifying the binocular images via convolutional neural network (CNN) in the proposed method. We divide the screen of the desktop computer into 2 x 3 and 6 x 9 blocks respectively, label the binocular images with their corresponding gazed block positions, train a convolutional neural network model to classify the eye images according to their labels, and estimate the gazed block through the CNN-based classification. The experimental results demonstrate that the proposed gaze estimation method based on binocular images can reach higher accuracy than those based on monocular images. And the proposed method shows its great potential in practical touch screen-based applications.