Evaluation of state-of-the-art pupil detection algorithms on remote eye images

Abstract

Eye movements are a powerful source of information as well as the most intuitive form of interaction. Although eye-tracking technology is still in its infancy, it offers the greatest potential for novel communication solutions and applications. Whereas head-mounted eye-trackers are widely used in research, several applications require most unintrusive eye tracking, ideally realized by means of a single, low-cost camera placed away from the subject. However, such remote devices usually provide low resolution images and pose several challenges to gaze position estimation. The key challenge in such a scenario is the robust detection of the pupil center in the recorded image. We evaluated eight state-of-the-art algorithms for pupil detection on three manually labeled data sets recorded in remote tracking scenarios. Among the evaluated algorithms, ElSe [6] proved to be the best performing approach on overall 3202 images from remote eye tracking, which include changing illumination, occlusion, head movements, and off-axial camera position. In addition, we contribute a new data set with 445 annotated images, recorded in a fixed setup with a low cost camera capable of using natural and infrared light.