Locating Corresponding Regions in Urban Environments

Abstract

This paper presents a novel approach to locate corresponding regions between two views in urban environments despite the presence of repetitive structures and widely separated views. First we extract hypotheses of building facades, each defined by a rectangular region. The inputs from each pair of regions in two images derive a projective transformation model. Extracted lines and points are used to evaluate the transformation model by voting on the correctness of the model. For each region, we find its corresponding region with the highest votes. Thereafter, we have a set of most corresponding pairs of regions. We take those frequently occurred transformation models (1-3) as correct models since many of the regions in the set come from the same plane, or else they are outliers. We redo the finding most corresponding regions step with the correct transformation models for those outliers with higher votes than the best one among each correct model group. Lastly, we choose the best pair of most corresponding regions with the highest votes among each correct model group. Experimental results show that our approach is effective and reliable in the case of viewpoint, orientation and scale changes.