Google will soon recognise locations just from looking at your holiday snaps
Google’s image recognition technology is improving so much it can recognise places from your photos
Google Photos is already pretty scary stuff. Upload a photo to it and it can categorise it based on its contents, identifying people, animals and landmarks – even food, plants and miscellaneous objects can’t hide from Google’s gaze.
Now, however, the Alphabet-owned company is working on improving this technology so it can identify the location where a photo was taken, even if that information isn’t present in the photo’s EXIF data. Spooky stuff.
Created by Google computer vision specialist Tobias Weyand, Google’s tech – dubbed PlaNet – can work out where a photo was taken by comparing it to a database of geo-tagged images available online. Speaking to MIT Technology Review, Weyland explained that the project came about from his team of engineers attempting to determine the location of photos from the pixels it comprised of. The team then divided the Earth into a grid of 26,000 squares of varying size depending on the number of available images of each location.
To teach the neural network to recognise locations, the team used its database of 126 million images. While all those images contain location data stored in the image EXIF, Weyand’s team showed PlaNet 91 million images without EXIF data to help it recognise places from just a photo.
In practice, PlaNet’s results sound a little underwhelming. Having used 2.3 million geo-tagged images from Flickr, PlaNet was only able to determine the photo’s country of origin with 28.4% accuracy – identifying the continent of origin 48% of the time.
However, as MIT Technology Review suggests, the technology already outperforms humans at the same task. In a game of Geoguessr, PlaNet trounced its human opponents: “In total, PlaNet won 28 of the 50 rounds with a median localisation error of 1131.7km, while the median human localisation error was 2320.75km,” said Weyand. “[This] small-scale experiment shows that PlaNet reaches superhuman performance at the task of geolocating Street View scenes.”
Given more time, images and feedback, PlaNet could become far better at recognising places from photos.
[Image: Tom Hall - Flickr]