Using AI To Identify Car Models In 50 Million Google Street Views Reveals A Wide Range Of Demographic Information

Google Street View is a great resource for taking a look at distant locations before travelling, or for visualizing a nearby address before driving there. But Street View images are much more than vivid versions of otherwise flat maps: they are slices of modern life, conveniently sorted by geolocation. That means they can provide all kinds of insights into how society operates, and what the differences are geographically. The tricky part is extracting that information. An article in the New York Times reports on how researchers at Stanford University have applied artificial intelligence (AI) techniques to 50 million Google Street View images taken in 200 US cities. Since analyzing images of people directly is hard and fraught with privacy concerns, the researchers concentrated on a proxy: cars. As an academic paper published by the Stanford team notes (pdf):

Ninety five percent of American households own automobiles, and as shown by prior work cars are a reflection of their owners’ characteristics providing significant personal information.

First, the AI system had to be trained to find cars in the Google Street Map images. That’s something that’s easy for humans to do, but hard for computers, while the next stage of the work — identifying car models — is much easier using AI. As another paper reporting on the research (pdf) explains:

the fine-grained object recognition task we perform here is one that few people could accomplish for even a handful of images. Differences between cars can be imperceptible to an untrained person; for instance, some car models can have subtle changes in tail lights (e.g., 2007 Honda Accord vs. 2008 Honda Accord) or grilles (e.g., 2001 Ford F-150 Supercrew LL vs. 2011 Ford F-150 Supercrew SVT). Nevertheless, our system is able to classify automobiles into one of 2,657 categories, taking 0.2 s per vehicle image to do so. While it classified the automobiles in 50 million images in 2 wk, a human expert, assuming 10 s per image, would take more than 15 y to perform the same task.

The difference between the two weeks taken by the AI software, and the 15 years a human would need, means that it is possible to analyze much larger data collections than before, and to extract new kinds of information. This is done by using existing datasets, for example the American Community Survey, which is performed by the US Census Bureau each year, to train the AI system to spot correlations between cars and demographics. The New York Times article lists some of the results that emerge from mining and analyzing the Google Street Map images, and adding in metadata from other sources:

The system was able to accurately predict income, race, education and voting patterns at the ZIP code and precinct level in cities across the country.

Car attributes (including miles-per-gallon ratings) found that the greenest city in America is Burlington, Vt., while Casper, Wyo., has the largest per-capita carbon footprint.

Chicago is the city with the highest level of income segregation, with large clusters of expensive and cheap cars in different neighborhoods; Jacksonville, Fla., is the least segregated by income.

New York is the city with the most expensive cars. El Paso has the highest percentage of Hummers. San Francisco has the highest percentage of foreign cars.

The researchers point out that the rise of self-driving cars with on-board cameras will produce even more street images that could be fed into AI systems for analysis. They also note that walking around a neighborhood with a camera — for example, in a smartphone — would allow image data to be gathered very simply and cheaply. And as AI systems become more powerful, it will be possible to extract even more demographic information from apparently innocuous street views. Although that may be good news for academic researchers, datamining offline activities clearly creates new privacy problems at a time when people are already worried about what can be gleaned from datamining their online activities.

Follow me @glynmoody on Twitter or, and +glynmoody on Google+

Permalink | Comments | Email This Story