Evaluation of a New Kernel-Based Classifier in Eye Pupil Detection

Accurate pupil location is paramount to applications such as gaze estimation, assistive technologies and several man-machine interfaces as the ones found in smartphones and VR applications. We introduce a new classifier stemmed from the Inner Product Detector and investigate its features on the challenging task of pupil localization. IPD (Inner Product Detector) is a classifier with high potential in facial landmarks detection. It is robust to variations in the desired pattern while maintaining good generalization and computational efficiency. However, one possible limitation is its linear behavior, which could be overcome by aggregating non-linear techniques, such as kernel methods. Although kernel classifiers have been exhaustively studied in the past two decades, it was not analyzed or applied with IPD, yet. The proposed KIPD achieves in the worst case an accuracy of 97.41% on the BioID dataset and 93.71% in LFPW dataset both at 10% of the interocular distance. In this paper the KIPD is compared to the state of the art methods, including the ones using deep learning, being competitive in terms of accuracy as well as computational complexity.