The applicability of cycle GANs for pupil and eyelid segmentation, data generation and image refinement

Eye tracking is increasingly influencing scientific areas such as psychology, cognitive science, and human-computer interaction. Many eye trackers output the gaze location and the pupil center. However, other valuable information can also be extracted from the eyelids, such as the fatigue of a person. We evaluated Generative Adversarial Networks (GAN) for eyelid and pupil area segmentation, data generation, and image refinement. While the segmentation GAN performs the desired task, the others serve as supportive Networks. The trained data generation GAN does not require simulated data to increase the dataset, it simply uses existing data and creates subsets. The purpose of the refinement GAN, in contrast, is to simplify manual annotation by removing noise and occlusion in an image without changing the eye structure and pupil position. In addition 100,000 pupil and eyelid segmentations are made publicly available for images from the labeled pupils in the wild data set (DOWNLOAD). These will support further research in this area.