The algorithm, the team said, has the potential to help architects, urban planners, and municipal governments make informed decisions about city structures and help rebuild cities after disasters.
"The software gives insights that are hard to obtain simply by analyzing raw geographic data," said Niloy Mitra, a computer science professor at the University of California, Los Angeles. "We hope this will help stakeholders design buildings and cities that are harmonious with the existing style."
The team focused on the architectural styles of Paris, identifying distinct neighborhoods and the stylistic essence that makes the city recognizable the world over.
The researchers gathered a database of 23,000 building polygons from the open-source OpenStreetMap project and manually labeled 2,000 of them to train a machine-learning model. The training data was created by breaking each façade down into simple line segments and then labeled by the experts.
Using this data, the team created a tool called "StyleFormer," a building shape generation model. StyleFormer enables the creation of a new façade in a particular architectural style or the modification of an existing façade according to a target architectural style.
"StyleFormer allows stakeholders to perform counterfactual analysis—they can envision 'what if' scenarios. For example, they can modify a building's façade to see whether the modification better aligns with the area's architectural style or evaluate whether a building's façade might look more appealing if altered in a certain way," Mitra said.
