Error Metric Using Correlation Between Binocular Corneal Images

The 3D point-of-gaze has been calculated using the binocular visual axes in the model-based eye gaze estimation. However, the 3D point-of-gaze is often shifted largely when the visual axis includes certain errors. The distance between the estimated point-of-gaze and the visual axis can be used as an error metric; however, horizontal errors cannot be evaluated, and an additional error metric is required to assess the reliability of the estimated point-of-gaze. Therefore, we propose a novel error metric using the correlation between binocular corneal images. We hypothesize that binocular corneal images extracted around the reflection of the estimated point-of-gaze exhibit a high correlation when the 3D point-of-gaze is correctly estimated. The two approaches were implemented using the corneal images extracted directly and generated based on binocular corneal imaging, and these approaches were evaluated through comparative experiments. We confirmed the feasibility of the proposed error metric and the effectiveness of binocular corneal imaging.