Monocular 3D gaze estimation using feature discretization and attention mechanism

Gaze estimation has become an important field of image and information processing. Estimating gaze from full-face images using convolutional neural network (CNN) has achieved fine accuracy. However, estimating gaze from eye images is very challenging due to the less information contained in eye images than in full-face images, and it's still vital since eye-image-based methods have wider applications. In this paper, we propose the discretization-gaze network (DGaze-Net) to optimize monocular three-dimensional (3D) gaze estimation accuracy by feature discretization and attention mechanism. The gaze predictor of DGaze-Net is optimized based on feature discretization. By discretizing the gaze angle into K bins, a classification constraint is added to the gaze predictor. In the gaze predictor, the gaze angle is pre-applied with a binned classification before regressing with the real gaze angle to improve gaze estimation accuracy. In addition, the attention mechanism is applied to the backbone to enhance the ability to extract eye features related to gaze. The proposed method is validated on three gaze datasets and achieves encouraging gaze estimation accuracy.