Buildings recovery employing Manhattan-world constraint

The structure from Motion (SfM) algorithm is an established method of shape recovery from a single video sequence. However, limitation of the method is the accuracy of the recovery due to poor texture of a recovered plane or a large scale object. We propose a technique for improving the precision of the recovery by applying the Manhattan-world constraint to the SfM algorithm, which assumes that the buildings are composed of vertical and horizontal planes. We show its effectiveness by the experiments performed in a real-life environment.