Generalizing over aspect and location for rooftop detection

Abstract

We present the results of an empirical study inwhich we evaluated cost-sensitive learning algorithmson a rooftop detection task, which is one level of processing in a building detection system. Specifically, weinvestigated how well machine learning methods generalized to unseen images that differed in location andin aspect. For the purpose of comparison, we includedin our evaluation a handcrafted linear classifier, whichis the selection heuristic currently used in the buildingdetection system. ROC analysis showed that, whengeneralizing to unseen images that differed in locationand aspect, a naive Bayesian classifier outperformednearest neighbor and the handcrafted solution.