Google is working on a new system that can identify the location of an image just by looking at it. Developed by Tobias Weyand, a computer vision specialist at Google, the goal of PlaNet is a computer system which can see a photo and instantly identify its location. According to his research, the result is a system “capable of localizing a large variety of photos. Besides landmark buildings and street scenes, PlaNet can often predict the location of nature scenes like mountains, waterfalls or beaches with surprising accuracy.”

PlaNet’s abilities were recently put to the test against human users. In an online game that displays random images from Google Street View, PlaNet was able to pinpoint an image’s location on a map of the world with greater accuracy than a well-travelled human player 56% of the time. That may seem like a low number, but consider the depth of knowledge and visual cues that a seasoned traveler has remembered. Additionally, many of the images lacked location markers, including pictures taken in nature or indoors. If you’re still in doubt, the game is available at www.geoguessr.com and it’s more difficult than it sounds. 

PlaNet is an artificial neural network, a digital system based on the structure of biological neural networks, in other words, human brains. Information that is input into such a system affects its performance. Mimicking a real brain it makes changes to how it operates based on input and the success of its output. In this way, it can learn new things. In the case of the PlaNet neural network, Weyand and his fellow researchers partitioned the planet into a grid of about 26,000 squares. They ten created a database with more than 100 million images with a known location to use as a reference for its visual recognition capabilities.

Whereas a typical human user, even a seasoned traveler, would rely on a combination of instinct and visual cues to identify a picture’s location, a system such as PlaNet can visually identify weather patterns, types of plant and animal life, and architectural styles. PlaNet is also able to analyze series of photos by combining its functionality with a “short-term memory architecture,” collectively studying images in a series, such as in a photo album, to correlate their location. 

In a test using 2.3 million images, PlaNet determined the country of origin 28.4 percent of the time and the continent at least half of the time. As the system develops these results will improve. It’s interesting to consider how a system such as PlaNet will be used in the future. Perhaps it could be used by curators to organize massive archives of photographs, by marketers looking to target users based on location, or for surveillance purposes.