Gaze Estimation Based on Attention Mechanism Combined With Temporal Network

Abstract

Due to the continuous and dynamic nature of gaze estimation, the true gaze point at each moment is closely related to the previous moment. Simply detecting individual frames of facial images cannot yield accurate gaze information. In current CNN-based gaze estimation methods, the effective utilization of eye movement temporal information and the ability to capture global relationships in the feature extraction process remain problematic. Addressing these concerns, this paper proposes a novel gaze estimation framework, named FE-net, which incorporates a temporal network. This framework introduces channel attention modules and self-attention modules, enhancing the comprehensive utilization of extracted features and reinforcing the contribution of valuable regions to gaze estimation. We further integrate an RNN structure to learn the temporal dynamics of eye movement processes, significantly improving gaze direction prediction accuracy. This framework predicts the gaze directions of left and right eyes separately using monocular and facial features and computes the overall gaze direction. FE-net achieves state-of-the-art accuracy of 3.19° and 3.16° on the EVE dataset and the MPIIFaceGaze dataset, respectively.