Google's New AI Can Guess Where Your Photos Were Taken

You can’t hide from Big Brother

If you use your smartphone to take pictures, there’s a good chance they’re tagged with a geographical location. Unless you’ve chosen to turn the setting off, the location is saved with the picture, and you can choose to share it alongside the photo when you post it on social media.

You might think that the location feature gathers too much information about you and that you’d like to keep your whereabouts private, but at least in that case you have a choice. Google’s new artificial intelligence (AI) technology can figure out where your pictures were taken, whether you like it or not.

SEE ALSO: New Satellites Will Detect Your Face and Phone from Space

Computer vision specialist Tobias Weyand and his colleagues at Google created a deep-learning program called PlaNet that searches for visual elements — landmarks, food, pets, colors, etc. — in pictures and identifies the location where they were taken.

To train the AI, they fed it 91 million geolocated photos to analyze based on a grid that divides the world into 26,000 squares of different sizes based on how common pictures from that region are. Areas like the Arctic and the middle of oceans were omitted as very few pictures are captured there.

They validated it with another 31 million images and then put it to the test with 2.3 million images from Flickr. It was scary good!

Granted, it wasn’t perfect, but “PlaNet is able to localize 3.6 percent of the images at street-level accuracy and 10.1 percent at city-level accuracy,” Weyand and colleagues write in their research paper. That might not sound great, but when you bring the desired precision level up to the size of a country or continent, it was accurate 28.4 and 48 percent of the time respectively.

Still doubtful? See how accurately you can guess the location of pictures by playing Geoguessr — it’s a lot harder than it sounds! Weyand and his team had PlaNet play the game with astounding results.

“In total, PlaNet won 28 of the 50 rounds with a median localization error of 1131.7 km, while the median human localization error was 2320.75 km,” Weyand and colleagues say.

PlaNet has the advantage that is had “seen” more places around the world than a human could ever travel to. Thus, it can recognize subtle cues that people aren’t likely to pick up on. Furthermore, since PlaNet is an artificial neural network, it can continue to learn and improve over time. The more geotagged images it sees, the better it will get.