We do not mean to construct a reasonably accurate copy of Manhattan. If you could, you could license your model to lots of interested parties for hundreds of thousands of dollars. Here is an interesting NYT article (Sat Nov 22, 1997, page B1) about real life models of Manhattan.
We said "reasonably accurate model" because a no model would be perfectly accurate for long! A city is a living thing that changes daily. But in fact our goal is even less ambitious. We want to construct a city that is recognizably Manhattan (and not Boston, L.A., San Francisco, or even Brooklyn). What do we mean?
What makes Manhattan "Manhattan"? We need to appreciate this in order to design our tools. An obvious aspect is the famous "Manhattan grid-like blocks". But as we all know, this geometry breaks down below 14 street, and even uptown, it is upset by features such as Broadway. It is just as important that our model displays such exceptions to the rule. For instance, no skyline of Manhattan is acceptable without the Empire State Building in midtown and the World Trade Center in downtown. In other words, recognizable landmarks need to be explicitly modeled.
Most most tall buildings in Manhattan are in the financial district and in central midtown area. We can generate a model with this property by first breaking down the city into neighborhoods, each with their own characteristics. Each neighborhood can be given statistical parameters such as the average height of buildings. Notice that this means that the individual buildings, including their heights will be generated randomly!
An important aspect of our use of statistical properties is its top-down propagation properties: neighborhoods are broken down into blocks. When we specify a statistical parameter for a neighborhood, this parameter will in turn automatically generate similar parameters for each block in the neighborhood. E.g., if neighborhood has an average building height of 60 feet, and (say) it has 4 blocks, then our program may make the average heighs of these blocks 100, 60, 40 and 40 respectively. Next, in the block with average height 100 feet, the distribution of actual building heights might range from 180 feet to 25 feet. Thus global parameters are propagated from neighborhoods to blocks to buildings.
Second order effects can be considered as well: the standard deviation of building heights and their distribution can be modeled. The heights of blocks in adjacent neighborhoods might have an influence on each other.
For our purposes, we view Manhattan as the union of 10 super-neighborhoods:
Here is an xfig file to illustrate this division. This is just to guide our actual modeling parameters - in the current formulation, we can specify neigbhorhoods but not super-neighborhoods.
Note that although our immediate goal is Manhattan, once our tools are in place, presumably we could use different parameters to generate models with the characteristics of other cities. At the end, we discuss some possible extensions.
Another critical remark about our model is the extensive use of textures. This is the only way to generate believable scenes of any complexity!