Democratizing eye-tracking? Appearance-based gaze estimation with improved attention branch

Appearance-based gaze estimation in 2-dimensional screen coordinates–the prediction of the users’ gaze from webcam footage–cannot yet compete in accuracy with infrared (IR) eye trackers. Yet by circumventing the constraints of requiring dedicated hardware, it shows great potential in many technological industries, as evidenced by some readily available commercial solutions, bringing democratization of eye tracking closer to the people. We present Residual Appearance-based Gaze Estimation network (RAGE-net), a novel convolutional neural network for gaze estimation without need of calibration, utilizing a fraction of computational resources required by similar networks, while also achieving competitive accuracy. The angular error is measured as 4.08°in the MPIIFaceGaze dataset (Max Planck Institute for Informatics Faze Gaze) and 3.96°in the MPIIGaze dataset. The architecture’s principles, covered by a comprehensive ablation study, include an attention branch, residual learning, weight sharing between eye channels, batch normalization and an eye image input normalization pipeline that removes dependence on full face input. With RAGE-net, we conduct an applicability study for gaze estimation approaches of similar accuracy for interpreting on-screen gaze in praxis. Findings demonstrate low heatmap validity, with coarse heatmaps as potential adaptation to approximate IR eye tracking. The effects of environmental factors such as camera position, illumination, distance and glasses are analyzed in-depth. • Proposed network with residual learning, attention mechanism and reduced input. • Angular error of 4.08°achieved with reduced network complexity 9.79 GFLOPs. • Ablation study of design modules, eye input cropping and normalization is presented. • Applicability of heatmaps from appearance-based gaze estimations is analyzed. • Impact of environment on webcam gaze tracking accuracy is investigated.