A new program can now determine where your picture was taken even without geotagging technology. With this new neural network machine known as PlaNet, it can figure out the exact location of an image by solely identifying its pixels.

Like Us on Facebook

This new project is developed by Google computer vision specialist, Tobias Weyand who trained this new machine to have a better ability than humans in determining the location of images even if there are no visible clues of where it was captured. 

The PlaNet team developed this program by mapping out the world into a grid made of 26,000 squares. These squares possess different sizes which will depend on the number of photos taken in the areas within a specific square. Popular locations like tourist spots and big cities are usually composed of more numbers of smaller squares, where more remote locations have fewer yet larger squares.

This database now has 26 million geolocated pictures that are taken from the internet where the squares are related to where they were taken. Among all of these photos are 91 million that were utilized to train this complex network to determine the location of an image that is just determined by the image itself.

The remaining 34 million photos in the database was used to validate PlaNet to test its capabilities.This deep learning program analyzed 2.3 million photos from Flickr and was able to pinpoint the location of 3.6 percent of the images at street level accuracy and 10.1 percent of accuracy for cities. Apart from this, PlaNet also determined the country where the photo was taken among 28.4 percent of all images, and the continent at 48 percent accuracy.

The program was also compared to human ability as it was tested versus 10 individuals that have wide travelling experiences, by showing photos to PlaNet and the test subjects that are obtained from Google Street View.

After 50 rounds, PlaNet fared better by winning 28 of them where human guesses were off by an average of 2,320.75 kilometers as PlaNet was off by only 1,131.7 kilometers.

Researchers reveal PlaNet's secret to its location identifying abilities, where Weyand says that the advantage of a deep learning program is that it can see more places than humans that allowed it to learn from clues, compared to humans who are not able to recognize these clues.

PlaNet only takes up 377 MB of space which can also be used as an app for smartphones.