Gable Roof Description by Self-Avoiding Polygon

In this paper, we present a Self-Avoiding Polygon (SAP) model for describing and detecting complex gable rooftops from nadir-view aerial imagery. We demonstrate that a broad range of gable rooftop shapes can be summarized as self-avoiding polygons, whose vertices correspond to roof corners. The SAP model, defined over the joint space of all possible SAPs and images, combines the shape prior embedded in SAP and a set of appearance features (edge, color and texture) learned from training images. Given an observed image, the posterior probability of the SAP model measures how well each SAP fits the observed data. Our inference algorithm follows the MAP framework, i.e. detecting the best gable roof is equivalent to finding the optimal self-avoiding polygon on the image plain. Even though the entire state space of all SAPs is enormous, we find that by using A∗ search, commonly our algorithm can find the optimal solution in polynominal time. Experiments on a set of challenging image shows promising performance.