Google Unveils Neural Network Machine Vision Technology That Recognizes The Location Of Any Image

If you're a fan of travel, you're probably going to love what Google has been up to lately. As we've covered many times, Google is no stranger to machine-learning and developing complex and powerful neural networks of computer resources. Now, the company has decided to focus on geography as well as testing you, not only its own neural network.

Google's Tobias Weyand, with the help of other engineers, has developed a neutral network that helps pinpoint exactly where in the world a photograph was taken. For things like landmarks, this wouldn't be too complicated. But what about inside someone's home? Inside a shopping mall? On an ordinary road like the one below?

The image above was pulled straight from the "GeoGuessr" game that has spawned out of this effort at Google, and it's a perfect example of a difficult guess. It's also an example of a photo Google's neural network couldn't make an accurate guess for, because there's just so little real information to go on. If a commercial vehicle was in the scene, or even a road sign, both humans and Google's network alike would have an easier time figuring it out.

To make the neural network work, areas of the world were split into 26,000 squares of varying complexity. The focus was capturing the world's most populated areas, so there is much of the Earth that is left untouched. That's probably just fine, as those remote areas could be even more difficult to guess--nature simply looks like nature in many parts of the world, with few varied characteristics to help guide you to the correct guess.

To beef up the database, a staggering 126 million geolocated images were scraped from the Web; 91 million of these were used to kick off the network, while the remaining images were used for validation.

All told, this neural network has been able to pinpoint 3.6 percent of the images with street-level accuracy, while 10.1 percent were city-level. Further, 28.4 percent of the time, country guesses would be correct, and for 48 percent, continent guesses were on the mark. Clearly, these results are not accurate enough for some use case beyond a casual game like this, but it's impressive nonetheless, and it will only continue to get better.

Now it's your turn to show how smart you are. You can hit up page here to play the "GeoGuessr" game, where you are dropped in randomly around the globe and have to make a dart throw guess on the world map about where it is. Expect to see many images like the one at the top of this post (that's east of the South African town of Theunissen, for the record). Once you are done, you can play more specific versions of the game, which may be a bit easier, or perhaps even more difficult.