Monocular Gaze Estimation Network based on Mixed Attention

Abstract

As an indicator of human attention, gaze is useful to diagnose mental state and predict human intention. Binocular fixation difference is the main standard to evaluate attention, which needs to be obtained by calculating the fixation points of both eyes. In recent years, appearance-based gaze estimation technology has been used in human attention diagnosis. However, most of the existing appearance-based gaze estimation methods only predict the gaze convergence point of both eyes. In this paper, we propose a Monocular Gaze Estimation Network based on Mixed Attention (MGE-Net). Through Convolutional Neural Networks(CNNs), we utilize the monocular feature, face feature, and monocular position information to predict the gaze point of each eye. Finally, the superior performance of MGE-Net is proved by experiments on the MPIIFaceGaze and GazeCapture dataset.