Multiple rectangle model for buildings segmentation and 3D scene reconstruction

This paper introduces a method for automatic extraction of buildings in aerial images. We first present a method based on rectangular buildings, which are the most common constructions. After a rough segmentation, we estimate a criterion of similarity of each region with the best matching rectangle. For buildings of complex shapes, we introduce an iterative way to divide a region in order to optimize its approximation by a set of rectangles. We use a parametric deformable model for refining rectangle size and positions. These rectangles are then used to enhance a 3D realistic reconstruction of the scene including building models.