Google has made an announcement that it has succeeded in developing, or one should say that the company has trained an all-new neural network termed as PlaNet. It won’t enhance video thumbnails but instead can know the location of any image without any extra information.
The PlaNet system can tell where an image has been taken and generally specifies it to a continent, though in a few particular cases it can even remarkably determine a city or even the exact street name on the basis of the origination of a given image.
It succeeded in doing so by analyzing pixels of an uploaded photo and looking for patterns on the Internet.
Obviously, PlaNet has lot of loopholes, including the localization process which provides street-level exact results can just be undertaken in 3.6% of cases, while only roughly every tenth picture can be narrowed down to the city when the image was clicked.
The PlaNet is much better at identifying the nation of origin of an uploaded image, having 28.4% accuracy, and can guess the shown continent in nearly every other case with a 48% success rate.
Though the network has been consistently evolving and the findings aren’t perfect, they’re still possibly quite correct for the present version as PlaNet has routed more than 2.3 million geotagged pictures posted on Flickr, which is among the most popular image-based social networks.
A group of scientists, engineers, and programmers headed by a ‘computer vision specialist’ Tobias Weyland has developed PlaNet.
The practical work on the project began when developers created roughly 26,000 geographical squares of different sizes which were drawn on the basis of the amount of pictures captured within their borders.
Thereafter, it analyzed more than 90 million pictures with unverified location data and subsequently put it in corresponding squares on the digital world map, after which an extra 34 million images with verified location data were used for the process of validation.