There are lots of ways to figure out where a photo was taken. From architecture to wildlife to clothing styles, we've been conditioned to recognize certain clues that might indicate the location of a scene. Now Google has developed a new A.I. system that can outperform humans in this area using only the visual information in a picture.
According to MIT Technology Review, the new neural network, dubbed PlaNet, was fed 2.3 million images from Flickr to test its capabilities. By looking at the pixels in each image, the system was able to determine the picture's country of origin 28.4 percent of the time and the continent with a success rate of 48 percent.
Instead of using GPS data, the software bases its guesses off a vast database of geotagged images collected from the Internet. It's even able to discern the location of images with no obvious clues, like those taken of objects indoors, by comparing them to other photos in the same album.
The team behind the project says this database gives PlaNet a leg up over its human competitors, because it has seen and gathered data from more places on Earth than one person could ever possibly visit. This idea was supported when the program went head-to-head against 10 well-traveled individuals to see who could recognize the most locations. The A.I. beat out the human team in 28 of the 50 rounds. If you want see how your location recognition abilities stack up against PlaNet, you can play the Geoguessr game online.
This new software is an exciting development in Google's artificial intelligence technology, which is already able to generate automatic email responses and produce some insane-looking artwork. And what's even more impressive about PlaNet is that it works without taking up too much memory. The program requires just 337 megabytes to run, which is small enough to fit on a smartphone.
[h/t MIT Technology Review]