Wondering where a fabulous picture on the internet was taken at – we’ve all been there. Guessing the location is quite hard if the image is not geotagged.
It seemed someone at Google had this problem, too, so the company decided to take the image search to the next level.
Google built a neural network called PlaNet, loaded with more than 90 million location-tagged images. When you search for the location of a picture – PlaNet scans it at pixel-level matches its contents to the already existing database.
The first batch of trials proved promising – PlaNet guessed the country with a 28.4% success rate and the continent – with a 48% success rate. PlaNet was also able to recognize 3.6% of times images at street level accuracy, and 10% at city-level.
Sure, those numbers hardly sound impressive, but they, in fact, do when you consider this is just the beginning.
Hopefully, we’ll see PlaNet up and working sooner rather than later. The service supposedly takes 377MB of memory, which means it can easily work even off a smartphone.