A simple but effective appearance-based gaze estimation method from massive synthetic eye images

Abstract

A novel method for appearance-based gaze estimation from massive synthetic eye images is proposed in this paper. This method is a combination of neighbor selection and gaze local regression for gaze mapping. First, a simple cascaded method using multiple k-NN(k-Nearest Neighbor) classifier is employed to select neighbors in feature space joint head pose, pupil center and eye appearance. Second, PLSR (Partial Least Square Regression) is applied to seek for a direct correlation between image feature and gaze angle. Experimental results demonstrate that the proposed method achieves state-of-the-art accuracy below 1 degree for with-in subject gaze estimation on public synthesis eye image dataset.