Google has pointed its staggering computer expertise to an interesting task: determining where any photo was shot by having artificial intelligence (AI) examine the background. Using a learning neural network and 92 million images, Google created PlaNet, a system capable of quickly locating the geographic location for most photos. And it does so with an accuracy better than most humans.

In an interview with MIT Technology Review, Tobias Weyand, a computer vision specialist at Google, revealed that he and some of his friends have trained a deep-learning machine to be able to identify where a photo was taken and came up with the PlaNet computing network.

The project

To kick off the project, the team divided the world into grids consisting of 26,000 squares of varying sizes. The size of the square depends on how many images were taken in the place, which means the more popular the place is to tourists, the more images it has, the bigger the square.

When a photo is taken using a smartphone or digital camera it creates Exchangeable Image File Format (EXIF) metadata which contains information such as the GPS location of where the photo was taken. The team created a database of images taken from the web and using the EXIF data, the team was able to sort out where each photo falls into the grid.

Image courtesy of Google

The database consisted of 126 million images with 92 million of them used to teach a powerful neural network to work out the grid location using only the image itself. They used the remaining 34 million images to validate images in the data set. The ideas was, if you put an image into the network, it will be able to pinpoint where the photo was taken, or at least some candidate places of where the image was possibly taken. The end result is PlaNet.

Meet PlaNet

To test the accuracy of PlaNet, the team fed it 2.3 million geotagged images taken from Flickr (Yahoo, Inc.) to see if it was able to identify the places where the photo was taken. Weyand stated that “PlaNet is able to localize 3.6 percent of the images at street-level accuracy and 10.1 percent at city-level accuracy,” and is able to determine the country of origin in a further 28.4 percent of the photos and the continent in 48.0 percent of them.

The team also tested PlaNet against 10 well-traveled humans via an online game called GeoGuessr, which determined that the network was better at identifying places. Weyand stated that “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.” He also added that “[This] small-scale experiment shows that PlaNet reaches superhuman performance at the task of geolocating Street View scenes.”

Weyand explains that PlaNet did better than the travelers because it has seen more places than any single human has ever been. PlaNet is also able to identify the location of photos with missing location cues such as those taken indoors by looking through other images placed in the same album to check more specific images to be able to determine the location of where the photo was taken.

Despite the huge amount of data gathered for this project, it only uses about 377MB of memory. Which means the model can easily fit into a smartphone. Imagine having a superhuman neural network right in the palm of your hands?

If you want to test how well you can identify places based on photos or you just want to burn some time, check out the game on https://www.geoguessr.com/.

Photo by DariuszSankowski (Pixabay)