3D gaze information is important for scene-centric attention analysis but accurate estimation and analysis of 3D gaze in real-world environments remains challenging. We present a novel 3D gaze estimation method for monocular head-mounted eye trackers. In contrast to previous work, our method does not aim to infer 3D eyeball poses but directly maps 2D pupil positions to 3D gaze directions in scene camera coordinate space. We first provide a detailed discussion of the 3D gaze estimation task and summarize different methods, including our own. We then evaluate the performance of different 3D gaze estimation approaches using both simulated and real data. Through experimental validation, we demonstrate the effectiveness of our method in reducing parallax error, and we identify research challenges for the design of 3D calibration procedures.
View on arXiv