A System for Building Detection from Aerial Images

We describe a method for detecting rectilinear buildings and constructing their 3-D shape descriptions from a single aerial image of a general viewpoint. 2-D roof hypotheses are generated from linear features by perceptual grouping. Good hypotheses are selected and then verified by computing wall and shadow evidence for them, which also provide the height information for the buildings. A 3-D reasoning process resolves conflicts among hypotheses in 3-D space. Results from several images can be integrated at a high level. An interactive system allows efficient editing of results by making use of the analysis performed by the automatic system; it also allows for some initial preparation of the data to improve results of the automatic system. Some results and their evaluation are included.