SEMI-AUTOMATIC 3D RECONSTRUCTION OF PIECEWISE PLANAR BUILDING MODELS FROM SINGLE IMAGE

This paper presents a novel algorithm that enables the semi-automatic reconstruction of man-made structures (e.g. buildings) into piecewise planar 3D models from a single image, allowing the models to be readily used for data acquisition in 3D GIS or in other virtual or augmented reality applications. Contrary to traditional labor intensive but accurate Single View Reconstruction (SVR) solutions that are based purely on geometric constraints, and recent fully automatic albeit low-accuracy SVR algorithms that are based on statistical inference, the presented method achieves a compromise between speed and accuracy, leading to less user input and acceptable visual effects compared to prior approaches. Most of the user input required in the presented approach is a line drawing that represents an outline of the building to be reconstructed. Using this input, the developed method takes advantage of a newly proposed Vanishing Point (VP) detection algorithm that can simultaneously estimate multiple VPs in an image. With those VPs, the normal direction of planes which are projected onto the image plane as polygons in the line drawing can be automatically calculated. Following this step, a linear system similar to traditional SVR solutions can be used to achieve 3D reconstruction. Experiments that demonstrate the efficacy and visual effects of the developed method are also described.