Parameterised models of Architecture

This is still work in progress, but the main goal is to automatically build 3D photorealistic models of architecture from image sequences. To do this our system has a predefined "Lego kit" of parts such as windows, pillars and doors which it tries to fit to the images.
Here are some VRML 1.0 models showing preliminary results. These are best viewed using CosmoPlayer 2.1, which is available here. Please be patient as the textures sometimes take a while to load!
Because our models are a collection of these "Lego kit" components, rather than just a set of texture-mapped triangles, we can automatically enhance them in interesting ways. For instance in the following models (VRML 2.0) the windows have all been made reflective and partially transparent, and texture from more visible windows has been mapped to those with less visibility:
For those without VRML browsers, here's a movie of the Trinity model (AVI, 2.2 MB), and the image sequence (436Kb) used to create it. A movie (MPEG, 2.2MB) of the Downing Library model is also available.
Associated with each of our components is a probability distribution which is based on architectural conventions and practical considerations. For instance, windows in Gothic buildings are likely to be tall and narrow, while Classical windows have a more square shape. Doors are likely to appear near ground level, and buttresses will stretch from the ground to the roof of the building. Using advice from architects, we try to define these pdfs in a way which is useful but not too restrictive on the types of building allowed. We can test our pdfs by sampling random buildings from them. An example of a random building sampled from our Classical and Gothic distributions is given below:
By defining these building components and probability distributions linking them, we can infer what the building probably looks like even in parts which are not visible in any image. For example the backs of the columns in the Downing Library model are correctly reconstructed even though they are occluded in all views. On a larger scale, we can estimate what the entire model looks like even though only two walls are viewed in the images, using symmetry and regularity constraints:
More information is available in my recent publications.