Virtual cities have many potential applications, ranging from computer games and movies to urban planning, education, commerce, and civil defense simulations. The authors describe a system based on image-based rendering, for automatically constructing virtual urban scenes from simultaneously acquired omni-directional images. In order to create a new view, the omni-directional images are sampled in a way similar to the sampling of a three-dimensional (3D) light field, and distortions are corrected using a simple technique, based on the assumption that frontal planes of buildings are orthogonal to the ground.
While most of the techniques described in this paper are known in the computer graphics and computer vision communities, the authors combine knowledge from both of these fields. They suggest several improvements, and they address the whole pipeline, from capturing and processing image data, to reconstructing, rendering, and augmenting urban scenes.
The paper is an easy read, and is interesting and entertaining. A small drawback is the lack of analysis and technical information. For example, how sensitive is the system to changes in the illumination (for example, from partial cloud cover)? What is the maximum possible speed of the vehicle during data acquisition? What is the resolution and number of images necessary in order to obtain a high-quality reconstruction? How long does it take to compute new views?
In summary, the paper is definitely worth reading, and it is particularly recommended as an overview paper for readers interested in virtual environments.