AI Learns to Identify Locations From Photos

If nothing is written on the back of an old family vacation photo, or no location is tagged to a digital image, artificial intelligence might be able to help. Three Stanford University graduate students have demonstrated AI's ability to look at a photo and tell where it was taken. That skill could be useful in positive ways, NPR reports, such as helping field biologists spot invasive plant species quickly in large areas or identifying where power lines are down. But it might also reveal data not intended to be public knowledge. "From a privacy point of view, your location can be a very sensitive set of information," said Jay Stanley, who studies technology for the American Civil Liberties Union.

The students who developed the Predicting Image Geolocations project trained the program by showing it images from Google Street View. Most of the time, it provides a pretty accurate answer when shown personal photos for the first time. PIGEON provides the correct country for about 95% of the queries, usually picking a spot within 25 miles of the actual photo location. The program can spot subtleties in the image that humans might miss, such as minor differences in weather, foliage, and soil. In a couple of tests, PIGEON came within 35 miles of the location of a Yellowstone National Park campsite and within a few blocks of a spot on a San Francisco street.

Stanley expressed concern that such a program could be used to track people's movements, including the federal government looking at personal photos to see if a person went to a country on a US watchlist. The current precaution of taking GPS location tags off photos being posted online might no longer offer protection, per NPR. Although the technology isn't perfect, it's just getting started. "The fact that this was done as a student project makes you wonder what could be done, by, for example, Google," Stanley said. (More artificial intelligence stories.)