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

Conventional appearance-based 3D gaze estimation methods generally use the roll of the head pose to represent the eyeball’s roll status by default. To reduce degrees of freedom of head poses, a normalization step was proposed to apply global transformations to images to make heads upright and eyelids horizontal. However, due to the ocular countering-rolling (OCR) response, the eyeball will rotate in the opposite direction when the head tilts to the side. After normalization, the eyeball will have an extra roll compared to the roll status of the eyeball when the head is not tilted. This roll from the OCR response causes a changed orientation of the eyeball in normalized eye images, which represents the roll status of the anatomical structure inside the eyeball and consequently affects gaze directions. Thus in this work, we propose a pipeline to regress the person-dependent anatomical variation as a calibration process with considering the OCR response, which can work with our proposed eye-image-based person-independent gaze estimator trained with real and synthetic eye images. The proposed method firstly brings the OCR response into the gaze estimation task, achieving better performances on the two benchmark datasets with fewer parameters under the real-time scenarios. With a replacement of a deeper network, compared to state-of-the-art methods, the proposed method is more efficient, achieving a). better average estimate (3.9% and 2.5% improvement), b). much better standard deviation (lower by 59.0% and 44.2%) and c). a much lower number of parameters (reduced by 88.0%).