A Hybrid Image and Model Based Approach to Photorealistic Building Reconstruction

Abstract

We introduce an algorithm that automatically aligns images with partial wireframe models to compute extrinsic camera parameters with respect to the model reference frame. Aligned imagery is fused with the model to incorporate high-resolution textures and to facilitate context sensitive image processing. The technique is designed to exploit the approximately planar structure commonly found in human-made environments such as building façades, interior walls, parking lots, and roadways.

Initially, images are processed to extract sets of lines that are consistent with a common vanishing point, called line pencils, that may be the result of planar structure in the perspective view. Pose is computed using a symbolic model matching technique that minimizes error between extracted line pencils and the projected wireframe model. Results from the three-dimensional city modeling domain show that the technique is capable of automatic alignment to a partial wireframe model. Calibrated imagery can be processed to segment planar structure based such as building façades, and to refine the wireframe model through context sensitive detection of façade details such as windows and doors.