Balconies are typically located on the third floor of buildings in Paris and on the second floor in Prague or London. Researchers from Carnegie Mellon University are developing software that uses geo-tagged images to visually characterize places. By the way, street lamp heads in Prague tend to be angular, while they tend to be rounder in Paris.
Researchers analyzed 120,000 images, including many from Google Street View, to identify trends and distinctive architectural elements in 12 cities (Paris, London, Prague, Barcelona, Milan, New York City, Boston, Philadelphia, San Francisco, San Paulo, Mexico City and Tokyo). The goal of the project was to create software that can be used as a kind of facial recognition for architecture. By identifying architecture that is both characteristic of a region and unique to that region, the software should be able to identify the location.
“Note that neither of these two requirements by itself is enough: sidewalks and cars occur frequently in Paris but are hardly discriminative, whereas the Eiffel Tower is very discriminative, but too rare to be useful,” the research paper reads.
By identifying subtle differences in architecture, the software is often able to identify the source location of an image, which researchers pointed out has several useful applications. Patterns of particular visual elements (such as balconies with cast iron railings) can be mapped to certain areas of a city where they usually appear (i.e., large thoroughfares in Paris). It is also possible to identify similarities and differences between cities or neighborhoods in order to cross-reference image retrieval. An area of Prague, for example, might look similar to a certain neighborhood in Paris, and the software can share that information.
“Finally,” the paper reads, “the proposed algorithm is not limited to geographic data, and might potentially be useful for discovering stylistic elements in other weakly supervised settings, e.g., 'What makes an Apple product?'”
To read the full research paper, What Makes Paris Look Like Paris, visit the university website.