Robustifying compressed projections for expressive iris localization

In this paper we describe a new system for eye center (pupil) localization. The patch centered on the eye is described by concatenations of integral and edge projections. Next, for dimensionality reduction, the Principal Component Analysis (PCA) technique is employed, while the discrimination among possible candidates is performed with a Bagged ensemble of Regression Trees (BRT) classifier. The accuracy is further increased by a least mean square (LMS) hyperboloid fit over BRT reported results. While the system is aimed at portrait of faces in various expression and gazes sights, we will shows that it does produce very accurate results under standard challenges. We successfully tested over two public databases proving robustness to various stresses like eye expression, gaze direction or eye occlusion while keeping the computation time low.