3D City Model Generation from Ground Images

In this paper, a 3D city model generation method is described. At first, the ground images around the area of interest are acquired using a camera mounted on a GPS device and digital compass, which provide the initial inaccurate pose of each image. An adjacency graph that spatially represents the adjacency between images or buildings is built to handle efficiently a huge number of unordered image sequences. A set of images with views of the same building is automatically grouped, and an optimization algorithm based on SFM corrects their poses. Finally, a method of global pose estimation is outlined that can register 3D isolated building models in a global coordinate system. We validate our approach with a set of experiments on some urban sites.