Kappa Angle Regression with Ocular Counter-Rolling Awareness for Gaze Estimation

Traditional 3D gaze estimation methods often use head pose roll to represent eyeball roll. To simplify head poses, a normalization step is employed, aligning heads and eyelids. Yet, due to the ocular countering-rolling (OCR) response, when the head tilts, the eyeball rotates oppositely. This results in additional roll post-normalization, affecting gaze direction. In our approach, we propose a pipeline that considers person-specific anatomical variations, addressing OCR effects, compatible with our eye-image-based, person-independent gaze estimator trained on real and synthetic images. This method integrates OCR responses into gaze estimation, outperforming benchmarks with fewer parameters in real-time scenarios. Compared to leading methods, it's more efficient, boasting improved average estimates (by 3.9% and 2.5%), significantly reduced standard deviation (by 59.0% and 44.2%), and a drastically lower parameter count (reduced by 88.0%).