The superhuman network can figure out the location of almost any photo using only the pixels it contains.
Humans are generally good at recognizing a famous building or landmark from an image like the Eiffel Tower, leaning tower of Pisa or the Roman Colosseum but the job becomes extremely difficult if the image does not contain a monument or any memorable place or if it is just an indoor image without clues or any other detail.
Buy Now: Sony PlaysStation VR In Stock Here
Now, Google has created a new freakish artificial intelligence machine that can identify places far more accurately than humans. This deep-learning machine can figure out the location of almost any photo using only the pixels it contains. Researchers have trained it to locate indoor images or even the pictures of pets and food through multiple visual clues such as weather patterns, vegetation and architectural styles.
The machine that they call “PlaNet” is developed by dividing the world into a grid of 26,000 squares. The squares are of varying sizes depending on the number of images taken in that location. Bigger cities have logically large squares where remote regions where photographs are less common have small squares. Researchers have ignored oceans and polar regions altogether, where very few photographs have been taken.
To train the machine, researchers created a dataset of 126 million images from the web. Then, they used 91 million images for training and 34 million for validation. Finally, researchers tested the powerful neural network in different ways to see how well it works.
For instance, researchers fed it 2.3 million geotagged images from Flicker to determine whether it will be able to identify the places accurately. PlaNet was able to locate streets at 3.6 percent accuracy while cities with 10.1 percent accuracy.
Next, it took on 10 well-travelled humans where all the participants were asked to pinpoint the location taken from Google Street View. The machine was able to outperform humans by a considerable margin. Overall, it won 28 of the 50 rounds with less localization error compared to humans.
“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.” Tobias Weyand, a computer vision specialist and lead researcher said in a statement.
Researchers further extended PlaNet to locate images that do not have location clues. The machine simply looked through the photo album to work out where they were taken.
Don't Miss: Nintendo Switch: Everything You Need To Know
“PlaNet is able to localize landscapes, locally typical objects and even plants and animals. Our experiments show that PlaNet far outperforms other methods for geolocation of generic photos and even reaches superhuman performance.” Study concludes.