Human Computer Interaction

The applicability of Cycle GANs for pupil and eyelid segmentation, datageneration and image refinement

We evaluated Generative Adversarial Networks(GAN) for eyelid and pupil area segmentation, data gener-ation, and image refinement. While the segmentation GANperforms the desired task, the others serve as supportiveNetworks. The trained data generation GAN does not re-quire simulated data to increase the dataset, it simply usesexisting data and creates subsets. The purpose of the re-finement GAN, in contrast, is to simplify manual annota-tion by removing noise and occlusion in an image withoutchanging the eye structure and pupil position. In addition100,000 pupil and eyelid segmentations are made publiclyavailable for images from the labeled pupils in the wild dataset.

Eye tracking is increasingly influencing scientific areassuch as psychology, cognitive science, and human-computerinteraction. Many eye trackers output the gaze location andthe pupil center. However, other valuable information canalso be extracted from the eyelids, such as the fatigue ofa person. We evaluated Generative Adversarial Networks(GAN) for eyelid and pupil area segmentation, data gener-ation, and image refinement. While the segmentation GANperforms the desired task, the others serve as supportiveNetworks. The trained data generation GAN does not re-quire simulated data to increase the dataset, it simply usesexisting data and creates subsets. The purpose of the re-finement GAN, in contrast, is to simplify manual annota-tion by removing noise and occlusion in an image withoutchanging the eye structure and pupil position. In addition100,000 pupil and eyelid segmentations are made publiclyavailable for images from the labeled pupils in the wild dataset.

Download

Source and Binaries

Publications