Eye Centre Localisation with Convolutional Neural Networks in High- and Low-Resolution Images

Abstract

Eye centre localisation is critical to eye tracking systems of various forms and with applications in variety of disciplines. An active eye tracking approach can achieve a high accuracy by leveraging active illumination to gain an enhanced contrast of the pupil to its neighbourhood area. While this approach is commonly adopted by commercial eye trackers, a dependency on IR lights can drastically increase system complexity and cost, and can limit its range of tracking, while reducing system usability. This paper investigates into a passive eye centre localisation approach, based on a single camera, utilising convolutional neural networks. A number of model architectures were experimented with, including the Inception-v3, NASNet, MobileNetV2, and EfficientNetV2. An accuracy of 99.34% with a 0.05 normalised error was achieved on the BioID dataset, which outperformed four other state-of-the-art methods in comparison. A means to further improve this performance on high-resolution data was proposed; and it was validated on a high-resolution dataset containing 12,381 one-megapixel images. When assessed in a typical eye tracking scenario, an average eye tracking error of 0.87% was reported, comparable to that of a much more expensive commercial eye tracker.