Google machine-learning system can figure out where you snapped a pic – without GPS
There’s more to taking pictures than just capturing pixels, and for photographers looking to put together a really robust gallery of all their snapshots over the years, meta data is key. We want to keep track of when we took a shot, where we took it, and maybe even who’s in it. That issue of “where” is easy enough when we’ve got location services enabled, but what if accidentally had it disabled for a while, or are importing a bunch of pics that somehow got their EXIF data from them along the way? Google’s been cooking up a new tool that attempts to roughly figure out where a photograph was taken – without using any GPS data at all.
Google’s team used a neural network to analyze 91 million images from points around the globe – pictures that still had embedded location data. After analyzing that set and looking for patterns, the system, now called PlaNet, is able to make an educated guess as to where a shot came from.
It’s far from perfect – figures right now have it guessing the correct city an image was taken in just over 10 percent of the time – but it’s only getting better. At larger scaled, accuracy is even better, naming the correct country 28.4 percent of the time, or correct continent in 48 percent of pics.
We don’t imagine PlaNet will replace GPS-based geotagging any time soon, but it’s still a fascinating experiment in machine learning, and one that could lead to some impressive, unexpected use cases as the technology matures.