3D Architectural Modeling From Multiple Images

With the introduction of digital city, the demand of large-scale three-dimensional (3D) model of a city has grown rapidly in recent years. As an important part of the urban scene, the creation of3D models of buildings is receiving more and more attention. Due to the convenience and low cost in data collection, image-based3D modeling of buildings has been a hot topic in computer vision and computer graphics fields.Aiming at3D architectural modeling from multiple images, this paper proposes a two-stage system which combines automatic reconstruction and interactive model-ing. Starting from several images of a building captured from multiple viewpoints, our system is able to construct a realistic3D model quickly and efficiently.In automatic reconstruction phase, this paper proposes a method to automatically construct the3D topological polygonal model of a building. First, we extract SIFT feature points for each image, and match the SIFT points for each image pair. Based on matched feature points, structure from motion (SfM) algorithm is applied to estimate the parameters of all cameras. Then multi-view stereo (MVS) algorithm is used to recover dense point clouds of the building from calibrated images. Next, we extract planar primitives from these discrete and unstructured point clouds. Then this paper presents a novel adjacency analysis method based on geodesic distance to obtain the adjacency relations of all planar primitives. For each face, its boundary is computed from the3D intersection lines with adjacent planes, its3D point clouds and corresponding2D image information. Finally, each face is mapped with texture to get a realistic3D model of the building.The3D model obtained in automatic reconstruction phase shows the main structure of the building, but due to the impact of image textures and occlusions, usually the3D model acquired from automatic reconstruction is not accurate, detailed and complete enough. Therefore this paper proposes interactive modeling to further edit the3D model obtained from automatic reconstruction, thus getting a finer3D model.In interactive modeling phase, we provide several simple operations for further processing of existing3D model. First, the user can select a plane to translate, rotate or scale, hence modifying plane parameters. Second, the user can edit the boundary of a plane on an image to snap it to its accurate boundary. And for the small indentations and protrusions in facades of buildings, we provide plane extrusion operation to construct them. In addition, our system offers other operations such as creating a new plane and seaming the gap between two planes, etc.By combining the strengths of automatic reconstruction and interactive modeling, the two-stage system proposed by this paper can construct a complete and detailed3D realistic model of a building quickly and efficiently. Finally, constructed3D models of several buildings are given, experimental results demonstrate the feasibility of our system.