Sure, you might be able to tell whether a city is a city by just looking at it. But can you train a machine to be even better than humans at recognizing them?
That's what a team of researchers sought to find in a paper titled "What Makes Paris Look Like Paris?" published in this December's issue of Communications of the ACM.
The researchers first gave 11 human subjects 100 random Google Street View photos, half of which were from Paris and half of which were from other cities, to see if they could pick out the city accurately. Even for nondescript scenes, subjects were right about 79 percent of the time. But the question was: how can we get machines to pick apart Paris' distinguishing features?
To do this, the researchers developed a program that took thousands of Google Street View photos and looked at randomly selected "patches" in those photos. While most patches were nondescript or frequently occurring (think trees, windows), flourishes like street signs, railings, or certain balcony styles tended to narrow down the results considerably. So the program focused on those distinctive, and filtered out the rest.
How useful could this be? According to the paper, we could end up using this create reference guides for various cities that could leverage huge amounts of existing data instead of sending out people to scout out locations for days. Imagine a TV series like Sense8, that takes place in a myriad of cities. Now imagine you could just get a computer to generate a "Paris-like" set without having to send out an entire crew to a location. The program could tell you, on short notice, what distinct features feature most frequently in any given city.
The program did run into a few problems during its trials, though. While it was able to point out several distinguishing features of distinct cities like Paris, US cities were much harder to pin down. You can imagine that would happen in places like New York City, where buildings tend to draw from a mish-mash of architectural styles. The program instead looked at things like cars and tunnels to characterize cities.
So can a computer be trained to recognize a city better than humans? It's hard to say, but one thing I can say for sure this: humans probably have just as hard a time as machines recognizing US cities without looking at major landmarks.