LAUSR

Precise eye center localization remains a very promising but challenging task, while its real-time performance constitutes a critical constraint in many human interaction applications. In this paper a new hybrid framework that combines the shape-based Modified Fast Radial Symmetry Transform (MFRST) and a Convolutional Neural Network (CNN), is introduced. The motivation of this work is to exploit the circularity of the iris to reduce the search space and consequently, the computational complexity of the fed CNN. Thus, the proposed hybrid scheme not only achieves real-time performance, but also increases substantially the localization accuracy by reducing the false detections of the MFRST. The experimental results that stemmed from the most challenging face databases demonstrated high accuracy, outperforming state of the art techniques even those that are based on end-to-end deep neural networks. To deal with unreliable data and provide valid evaluation, we manually annotated the FERET database, making the annotations publicly available. Moreover, the reduced computational time of the proposed scheme reveals that it can be incorporated in low-cost eye trackers, where the real-time performance is a basic prerequisite.