An analysis of accuracy requirements for automatic eyetracker recalibration at runtime

The initial calibration of an eye-tracker is a crucial step to provide accurate gaze data, often as a position on a screen. Issues influencing the calibration such as the user's pose can change while using the eye tracker. Hence, recalibration might often be necessary but at the expense of interrupting the user executing the working task. Monitoring interactions such as clicks on a target or detecting salient objects could provide recalibration points without deliberate user interaction. To gain insight into how accurate recalibration points must be localized to ensure that gaze estimation accuracy is improved, we conducted a user study and examined the effect of correct as well as erroneous localization of recalibration points. The results show that even a localization error of 1.2 degrees of visual angle induces an error of less than 0.5? to the estimated gaze position on screen. Our results indicate the necessary requirements any method automatically providing recalibration points has to fulfill.