Crowdsourcing pupil annotation datasets: boundary vs. center, what performs better?

Abstract

Pupil-related feature detection is one of the most common approaches used in the eye-tracking literature and practice. Validation and benchmarking of the detection algorithms relies on accurate ground-truth datasets, but creating of these is costly. Many approaches have been used to obtain human based annotations. A recent proposal to obtain these work-intensive data is through a crowdsourced registration of the pupil center, in which a large number of users provide a single click to indicate the pupil center [Gil de Gómez Pérez and Bednarik 2018a]. In this paper we compare the existing approach to a method based on multiple clicks on the boundary of the pupil region, in order to determine which approach provides better results. To compare both methods, a new data collection was performed over the same image database. Several metrics were applied in order to evaluate the accuracy of the two methods.