Robustness of Supervised Learning Based on Combined Centroids

Abstract

Recently, we proposed a novel sparse centroid-based supervised learning method, allowing to optimize a single centroid and its corresponding weights. The method is especially useful for localizing objects in images. Here, we extend the method to the task of joint localization of several objects in a 2D-image by means of combining several centroids. The novel approach, i.e. joint optimization of several centroids and a subsequent optimization of their weights, is illustrated on the task of localizing the mouth and both eyes in facial images. Because we are particularly interested in studying the robustness of the method to various modifications of the images, we evaluate the performance of the methods also over images artificially modified by additional noise, occlusion, changed illumination, or rotation. The novel centroid-based method is successful in the localization task, and the optimization turns out to ensure robustness with respect to the presence of noise or occlusion in the images. Moreover, combining the optimized centroids yields more robust results than a method using simple centroids with a highly robust correlation coefficient (with a high breakdown point).