RIT-Eyes: Realistically rendered eye images for eye-tracking applications

Abstract

Convolutional neural network-based solutions for video oculography require large quantities of accurately labeled eye images acquired under a wide range of image quality, surrounding environmental reflections, feature occlusion, and varying gaze orientations. Manually annotating such a dataset is challenging, time-consuming, and error-prone. To alleviate these limitations, this work introduces an improved eye image rendering pipeline designed in Blender. RIT-Eyes provides access to realistic eye imagery with error-free annotations in 2D and 3D which can be used for developing gaze estimation algorithms. Furthermore, RIT-Eyes is capable of generating novel temporal sequences with realistic blinks and mimicking eye and head movements derived from publicly available datasets.