3D Building Model Reconstruction Via Repetitive Pattern Analysis

The reconstruction of3D buildings is a topic of significant commercial and intellectual interest. Therefore it is no surprise that this research area has received significant attention in the past decades. Despite the high volume of existing work, there are many difficult unsolved problems which have high application value. Studies of these problems not only promote the development of reconstruction technology but also provide methods to other research areas.Several industrial and commercial applications benefit from the development of this research area. In the entertainment industry, the storyline of several movies and computer games takes place in cities. In order to make these cities believable at least some part of the models are obtained by3D building reconstruction. Digital map applications for mobile devices, cars, and desktop computers, e.g., Google Earth and Baidu Map, require2D and3D building models.In this paper we propose a novel method for multi-facade3D building model reconstruction from a single image. It is based on a fact that lots of urban buildings have horizontal repetitive structures on the surface. These repetitive structures include small objects like windows and balconies and complex structures like walls and textures. A group of repetitive structures which have the same shape and texture is a repetitive pattern. Although every structure belonging to the same pattern has the same shape and size in the real world, they look different in the image. It is because the image is generated by perspective projection. We detect a set of rectangular repetitive structures which are rectangular panels in the real world. By analyzing the relationship of these repetitive structures and the corresponding projections in the image, we calculate the orientations and positions of these panels. Further, we reconstruct the whole model of the building.In order to detect the repetitive pattern, we propose a novel scheme. First, we obtain the vanishing points in the image using an unsupervised method. There are two kinds of vanishing points:vertical vanishing points and horizontal vanishing points. Using these vanishing points, we obtain some basic perspective information and then adjust the image. In the next step, we divide the image into several vertical sections which correspond to rectangular panels in the real world. We calculate the similarity of these sections and group the similar ones. Then we merge the sections in a group with ones in another group if the corresponding sections are adjacent. The calculation of similarity is based on Mutual Information. We take into account not only similarity of the color but also similarity of the shape. Every merged group is a repetitive pattern which includes similar sections. Then we adjust every section in a group to maximize the similarity of the whole group. At last, we give a formula to measure the importance of every repetitive pattern, and select the optimal one as the input of3D construction stage.Our main contribution is reconstructing a multi-facade3D building model from a single image. While image-based modeling usually uses multiple images to reconstruct a building model, we give a method to complete this work using only a single image. Further, we have advanced the facade modeling technique to reconstruct multiple facades from a single image.