Kevin Willeford¹ (); ¹NOVA Southeastern University College of Optometry

Purpose. The limbus remains visible during all ocular rotations. Thus, the goal of our experiment was to evaluate whether gaze estimates based on limbal segments (LS) are as accurate as those derived from P-CR signals, and if so, to suggest that accurate gaze estimation is tractable for large eye movements. Methods. Twelve non-strabismic observers, head-stabilized with bite-bars, performed monocular fixations (5°- 30° in eccentricity along the 45° - 360° meridians) while P-CR signals (Eyelink 1000+) and ocular photographs were recorded. Limbal segments were extracted from photographs by first using a neural network to segment the iris and sclera and then defining the limbus as their common boundary. Next, an optimization routine transformed a reference limbus, obtained from photos captured in primary gaze, until the mean distance between points on the reference limbus and a given limbal segment was minimized. This approach produced transformation attributes (xSCALE, ySCALE, rotation, xTRANSLATION, yTRANSLATION) representing the unique shape and location of each observer’s limbi across all conditions. Two calibration functions, one using the transformation attributes and the other using P-CR signals, were created by using third-order polynomials to map each set of data to known fixation locations. Distance between predicted and actual fixations was used as a measure of accuracy. Results. P-CR tracking produced more accurate gaze estimates for right- [0.0124 +/- 0.0028 mm (P-CR) vs. 0.0204 +/- 0.0030 mm (LS), p = 0.04] but not left- [0.0238 +/- 0.0046 (P-CR) vs. 0.0199 +/- 0.0031 (LS), p > 0.05] eyes. This did not result from a difference in mapping quality (all R2 > 0.88) or from one method’s bias for greater error in a specific location. Conclusions. Limbal segments are potential proxies for measuring large eye movements, overcoming the recording range limitations inherent to P-CR trackers and capturing the eye’s full range of motion.

Acknowledgements: This research is supported by a President's Research Grant from Nova Southeastern University.