V-ir-Net: A Novel Neural Network for Pupil and Corneal Reflection Detection trained on Simulated Light Distributions

Deep learning has shown promise for gaze estimation in Virtual Reality (VR) and other head-mounted applications, but such models are hard to train due to lack of available data. Here we introduce a novel method to train neural networks for gaze estimation using synthetic images that model the light distributions captured in a P-CR setup. We tested our model on a dataset of real eye images from a VR setup, achieving 76% accuracy which is close to the state-of-the-art model which was trained on the dataset itself. The localization error for CRs was 1.56 pixels and 2.02 pixels for the pupil, which is on par with state-of-the-art. Our approach allowed inference on the whole dataset without sacrificing data for model training. Our method provides a cost-efficient and lightweight training alternative, eliminating the need for hand-labeled data. It offers flexible customization, e.g. adapting to different illuminator configurations, with minimal code changes.