According to reports, Google is developing software that can pinpoint where a photo was taken, on the pixel level.
On Wednesday, MIT Tech Review detailed the new project, led by Tobias Weyand, a computer vision specialist at Google. Weyand and his team have "trained a deep-learning machine to work out the location of almost any photo using only the pixels it contains." According to this report, the team started by dividing the world into a grid with more than 26,000 squares; bigger cities have a more "fine-grained" grid structure than remote regions and the team ignored areas like the Polar Regions and oceans as these are not common photography spots. Once this grid was in place, the team created a database using 126 million geotagged images from the web to teach the computer how to identify these locations. The neural network was taught using 91 million of these images and validated what it had learned with the remaining 34 million images.
After all this, the team put PlaNet through the gauntlet. To measure its accuracy, PlaNet was given 2.3 million geotagged images from Flickr to see if it could correctly determine the photos' location. "PlaNet is able to localize 3.6 percent of the images at street-level accuracy and 10.1 percent at city-level accuracy," said the team. PlaNet can correctly identify the country of origin in 28.4 percent of the images, and the correct continent in almost half. In tests against humans, PlaNet won 28 of the 50 rounds and has a lower median localization error than humans. The median human error was 2320.75 km while PlaNet's was less than half that at 1131.7 km. The team said, "[This] small-scale experiment shows that PlaNet reaches superhuman performance at the task of geolocating Street View scenes." Anyone can take on PlaNet's geolocating skills here.
Photo: © rvlsoft - 123FR.com