Can you guess where the above picture was taken? Google can. The Mountain View company is adding the final touches to a game changing new appliance. It is called PlaNet and it will be soon able to tell you where a photograph was taken by simply looking at it.
The Google PlaNet team trained the artificial intelligence to determine the locations where photographs were taken based on visual traces. The database is astonishing, it contains 126 million geolocated photos pulled from the Internet.
Google is working on it for years now. Regardless humans turned out to be quite good at recognizing locations, mostly because we can rely on our knowledge of the world’s differences, such as language on signs, different flora and fauna, architectural styles, etc, the Google team proved that the same task for regular humans is almost impossible and terribly difficult. They even tested this through GeoGuesser, a website where people have to guess a place in the world by looking at a Street View image and a few roads around. PlaNet seems to be accurate in this, and it will not need any extra data to figure out the location where the image was taken.
Thobias Weyand, Google specialist and leader of the PlaNet team, said: “We think PlaNet has an advantage over humans because it has seen many more places than any human can ever visit and has learned subtle cues of different scenes that are even hard for a well-traveled human to distinguish”.
This artificial intelligence was trained and now significantly outperforms humans. This machine is now able to identify the location of almost any photo using only the pixels it contains. To improve even more its geolocation abilities, for images that lack sufficient visual tips, the researchers trained the machine to do 'sequence location' – which means exploiting the way photos are generally taken in sequences and therefore the system can use entire photo albums.
The testing on this extra function was even larger. "For this task, we collected a dataset of 29.7 million public photo albums with geotags from Google+, which we split into 23.5 million training albums of 490 million images and 6.2 million testing albums of 126 million images" researchers said. Google PlaNet seems a total success until now. It’s very promising to see what Google’s AI is capable of, except from trying to beat the South-Korean Go champion!
Source: MIT. Image: Google