A Program That Can Identify Cities From One Image

For anyone who has seen the city of Paris, the image above is clearly from one of its streets. But, just what is it that distinguishes the unique Parisian style from other large cities around Europe? If you take a look at the image of a London street, below, the differences start to become more obvious… but even still, it would be easy to miss the subtler details of the two unique design languages. Enter the intriguing project What Makes Paris Look Like Paris.

Using the immense repository of geotagged imagery now available to us on the internet, a team at Carnegie Mellon University worked to create a system which would automatically find the unique visual elements of cities around the world; both the bold and the subtle. Using the city of Paris as an example, they used the system to comb through imagery from Google Street View and pull out elements like windows, balconies and street signs. They used a discriminative clustering approach, which selected for information such as the standard window shape and ignored such major landmarks as the Eiffel Tower and the Arc de Triomphe – as they only appear once.

By analyzing a sizable 250 million visual elements from 40,000 Street View images of Paris, London, New York, Barcelona and eight other cities, the system now has the knowledge to identify one of the cities using just one image. Not bad when you consider many people would have trouble distinguishing the two cities featured here. For a closer look at how the system distinquishes between locals, see the video at the bottom of this page, then see if you’re as good as the computer by taking their Paris/Non-Paris test (50 images are of Paris, 50 are of the 11 other cities). For more on the project, see graphics.cs.cmu.edu.

Elements of a Parisian street

Elements of a London street


Share this Story
Need help creating powerful branded content? Let Column Five hook you up.
Load More Related Articles

Facebook Comments

Get inspiration in your inbox. Sign up for our newsletter.