Google announced that it has managed to develop, or better said, train a brand new neural network called PlaNet. This one doesn’t improve video thumbnails, but instead is relatively capable of determining the location of an image without any additional information. This invention can therefore, pinpoint a place where a picture was taken and usually narrows it down to a continent, though in some particular cases it can even impressively deduce a city or even the exact street name from where a given image originates. It accomplishes this by analyzing pixels of an uploaded photo and looking for patterns on the Internet.
Naturally, PlaNet is far from perfect. The localization process which gives street-level accurate results can only be undertaken in 3.6 percent of cases, while just about every tenth image can be narrowed down to a particular city. As expected, it’s much better at specifying the country of origin of an uploaded picture with which it has 28.4% accuracy, and is able to guess the displayed continent in almost every other case (48% success rate). While the network is constantly evolving and the results aren’t perfect, they’re still probably pretty accurate for the current version as PlaNet has processed over 2.3 million geotagged photos uploaded to Flickr, one of the most popular image-based social networks.
PlaNet was developed by a team of scientists, engineers, and programmers led by a “computer vision specialist” Tobias Weyland. The practical work on the project has started after developers created around 26,000 geographical squares of various sizes which were drawn depending on the amount of photos taken within their borders. Over 90 million images with unverified location data were then analyzed and subsequently placed in corresponding squares on the digital world map, after which an additional 34 million pictures with verified location data were used for the validation process. Even more impressive is the fact that this solution only utilizes image Exif data, i.e. stored information about settings such as shutter speed, camera make and model that are used to take a particular image. This means that the latest edition of PlaNet is only 377 MB in size despite having around 126 million of indexed images in its database. The team behind this neural network even created a game called GeoGuessr powered by PlaNet which you can try out at the source link below. Apart from having fun guessing locations of images and competing against Google’s latest invention, users can also help develop the network further by playing. On average, the current version of PlaNet wins in 28 out of 50 rounds when competing against ten humans. This isn’t the first time Google has dabbled into neural networks; most recently, the Silicon Valley giant has managed to create a chatbot powered by similar technology.