Recognition of urban buildings with spatial consistency and a small-sized vocabulary tree

In this work, we address the problem of building recognition as a mobile application. Our approach exploits a small-sized vocabulary-tree of SIFT descriptors. Each SIFT descriptor in our dataset is saved along with its class label, its nearest neighbor from the vocabulary and the visual words corresponding to its spatial neighbors. To evaluate a new query image, we extract SIFT interest points and their descriptors and match it to a sub-list of descriptors that correspond to the same visual word. Then, as a verification step, we evaluate the spatial consistency. Finally, a voting scheme is used to decide which building category this image belongs to. The experimental results, obtained on two publicly available building datasets, show state of the art accuracy while ensuring reduced memory and computational requirements.