VarCity aims to drastically innovate 3D city modeling, moving it away from the traditional, bottom-up Lidar or structure-from-motion pipelines. Not only will large-scale city modeling be automated, but the results will look more realistic, will be more compact and will be semantically structured. Moreover, the dynamics of traffic flows will be added as animations (to safeguard anonymity) and videos of special events and landmarks can be retrieved with the 3D model as geographical context. These videos of flows and events are again automatically analysed and combined. For the combination of multiple event videos, a virtual editor will be implemented, ensuring compilations of high quality and providing an excellent overview whenever possible. The overall result will be an augmented, live Google street.The modeling is based on three types of input data: 1) dedicated mobile mapping imagery taken from several synchronized cameras mounted on a van, 2) images of landmarks and corresponding Wikipedia pages mined from public Internet repositories, and 3) the large amount of videos people can expect to stream from their mobile phones in the near future (given e.g. the upcoming LTE). With respect to the latter VarCity thus takes a proactive stand, but such data are easy to simulate. The other data types are already available to the project, through spin-off companies.For the modeling of the architectural structures, procedural methods will be used. Starting from the images, procedural models of the buildings are generated as instantiations of their architectural style, expressed in a style grammar. The key idea underlying much of the work is to let visual object class recognition and detection feed back into what traditionally are considered `low-level’ vision processes. This approach will enhance several aspects of the work, including the inverse procedural modeling, event detection, traffic flow analysis, inpainting, level-of-detail, super-resolution, etc.