Google. Either the most beneficial tech company on the planet, or masquerading as such long enough to take over the world. Either way, they’re amazing at making useful programs, the latest of which is “PlaNet,” a program that can geolocate any image in the world to its place of origin. Or at least that’s what designer Tobias Weyand, a computer vision specialist at Google, expects it to do in a short time if PlaNet continues to develop as quickly as it has.
If you were to try and guess the origin of a photo, you could make a pretty good effort by using certain visual cues: signs (and what languages they’re written in), architecture, vegetation, etc. You can actually try it out for yourself with this game, Geo Guessr.com, which challenges you to choose the correct place of origin of a google map location. Surprisingly, computers have not been so good at this, which is why PlaNet has been developed to make image location more accurate.
The PlaNet team got started by taking a look at the world map and ignoring less populated areas like polar and desert regions, remote and desolate lands, and large bodies of water. Next, they dedicated much of their attention to highly populated areas like big cities. Then the team took 126 million geolocated images that use Exif location data from the web to pinpoint exactly where those images were taken. 91 million of them were used to teach the program where images were taken, and 34 million of them were used to test the network and see how well it worked.
It worked. Quite well.
Taking a sample of 2.3 million geotagged images from Flickr, PlaNet could localize images to different degrees of accuracy. Of the 2.3 million, it could localize 3.6 percent–or 82,800–down to street level accuracy. It could localize 10.1 percent–or 232,300–to city level accuracy. It could localize 28.4 percent–or 653,200–to the country of origin. And it could localize nearly 50 percent–or 1,104,000–to the continent.
Where humans are adept at using all of the visual cues mentioned earlier for geolocating images, PlaNet uses even more, thanks to its ability to connect images by relevance. PlaNet can use photo albums and specific items shown in those albums’ to piece together the likelihood of a location. For example, perhaps there is a specific brand of snack food found only in the southern states of the U.S.: PlaNet can learn this from album data and apply it to future searches. Later, when it’s given an image that features that same brand of snack food, it can use that data as a factor in determining the image’s location.
How could this be implemented in the future? Weyland says that the current model only takes up 377 MB, small enough to fit on a smartphone. Perhaps this will be a new program Google releases in the future, which could only grow more accurate if millions of users continue to upload geotagged photos into it. Maybe Google is plotting to take over the world and PlaNet will now be able to track down its human victims with uncanny accuracy. What do you think? Let us know in the comments below and tell us how excited or terrified you are about Google’s impressive program.
Source: MIT Technology Review