You may already be familiar with Google’s popular reverse image search, where users can reverse search an image using the photo’s online URL or by uploading the photo itself to Google’s image search engine. After reverse searching an image, Google will return search results that show which websites the photo appears on and the possible location where the photo was taken.
Now, Google has figured out a way to find the location of almost any photo that is taken, even if the photo contains no geotag information in it. Google’s new artificial intelligence program called PlaNet can find the location where a photo was taken with up to 48% accuracy. The amazing thing is that PlaNet can identify the location where a photo was taken even if there are no famous or recognizable landmarks in the picture.
If you have ever played a round of Geoguessr, then you know how difficult it is to guess locations from random images. Geoguessr [ www.geoguessr.com ] is an online game where people try to guess the location of a randomly chosen Google Street View image.
Furthermore, PlaNet can even estimate the location of photos that were taken indoors or photos that contain people and pets.
PlaNet uses a neural network computer program to estimate image locations. The more images the neural network processes, the more it learns and the better it gets at estimating the geographic location for a given photo. Google’s team “taught” PlaNet by processing over 90 million Google Street View images, along with location information, from various places on Earth. The team then tested PlaNet’s accuracy by running 34 million public photos through the program that were pulled from Google+ and Flickr accounts.
PlaNet compares patterns and pixels in photos with the Google Street View images and locations in its database to estimate the locale for an image. The program uses satellite images and landscape patterns to estimate locations in photos of remote and rural areas.
PlaNet is able to outperform people in recognizing where a picture was taken. Google tested PlaNet’s ability to determine a picture’s location by pitting the program against 10 highly-traveled people in games on Geoguessr.
Amazingly, PlaNet won 28 out of 50 GeoGuessr games against its human competitors. In addition, the program’s median location error was only 1131.70 kilometers; whereas the human’s median location error was much higher at 2320.75 kilometers. This experiment shows that PlaNet can significantly outperform people in estimating photo locations even though its neural network program is still in the very early stages of “learning.”
As more images are processed through PlaNet’s neural network, the program’s location accuracy rates will get better and better. In addition, computers can store and process an almost unlimited amount of visual and location information to use for estimating a given photo’s geographic locale; whereas humans are unable to do the same.
So far, PlaNet’s photo location accuracy rates for various geographic levels are as follows:
- 3.6% at the street level
- 10.1% at the city level
- 28.4% at the country level
- 48% at the continent level
Google hasn’t said how, or if, their PlaNet program will be used, or whether or not it will ever be available to the public. However, the PlaNet program only needs 377 MB of memory, so it could easily run on a smartphone or mobile device for on-the-go location identification.
An image location program like PlaNet could have real-world applications like helping to search for people or to find missing persons pictured in recent photographs.