We introduce an accurate eye-tracking method that exploits deflectometric information and uses deep learning to reconstruct the gaze direction. We demonstrate real world experiments with evaluated gaze errors below 1°.
History
Preprint ID
111786
Highlighter Commentary
This study proposes a novel eye-tracking approach that uses dense deflectometric data and significantly improves accuracy compared to traditional methods. By training a Convolutional Neural Network model to predict gaze direction from deflectometry correspondence maps, the method achieves a gaze error of less than 1° and requires only a single camera and screen. Real-world experiments confirm its effectiveness, while future research aims to evaluate its performance on human eyes and explore further enhancements.
-- Mousa Moradi, Ph.D., Biomedical Engineering, UMASS Amherst