Robust classification and semi-supervised object localization with Gaussian processes

Abstract

Traditionally, object recognition systems are trained with images that may contain a large amount of background clutter. One way to train the classifier more robustly is to limit training images to their object regions. For this purpose we present a semi-supervised approach that determines object regions in a completely automatic manner and only requires global labels of training images. We formulate the problem as a kernel hyperparameter optimization task and utilize the Gaussian process framework. To perform the computations efficiently we present techniques reducing the necessary time effort from cubically to quadratically for essential parts of the computations. The presented approach is evaluated and compared on two well-known and publicly available datasets showing the benefit of our approach. © 2011 Springer-Verlag.