Combining Face with Face-Part Detectors under Gaussian Assumption

This paper addresses a simple and effective approach of face and face-part classifier fusion under Gaussian assumption, which is able to process heterogeneous visible wavelength (VW) and near infrared (NIR) image data. Evaluations using existing and publicly available AdaBoost-based individual classifiers on the recently released CASIA-V4 iris distance database of close-up portrait images as well as on YaleB indicate, that (1) single classifiers are largely affected by the type of training data, especially for NIR and VW data, and therefore prone to errors, (2) by combining individual classifiers a more robust classifier is obtained, (3) processing time overhead is negligible, if individual classifiers exhibit a low false positive rate, and (4) the proposed fusion approach is not only able to reduce false positives, but also false negative detections.