Here's a scene you didn't see in Oscar-nominated film "Zero Dark Thirty": Osama bin Laden is ranting and waving his AK-47 automatic rifle as he denounces the great Satan in a video broadcast on Al Jazeera. The video is recorded by U.S. intelligence and analyzed by a new technology that can pinpoint the location where a video was shot with an accuracy of 10 meters, and the orders go out to the drones. You know what comes next.
That didn't happen, of course. But a new technology currently cooking in a lab in Berkeley, Calif., could make that fantasy a reality. And we may regret it. Researchers at the Unversity of California's International Computer Science Institute are building a location-centric database by analyzing videos downloaded from the Web. By comparing a video whose provenance is unknown to the database of known videos, it can then make an informed guess about where the video was shot. And thus where anyone in that shot was when.
How video-location detection works
You've probably used speech recognition software that requires training. The software learns by making guesses about what you're saying, then comparing it to the corrections you make in the text. Pretty soon it knows that a Bostonian saying "Cuber" means Cuba and not cucumber. The video location software learns in a similar manner.
Jaeyoung Choi, the lead researcher on the project, downloaded thousands of videos from Flickr that contain embedded geographical information. That data may include location tags (aka geotags), visual cues such as textures and colors, time stamps, and sounds such as birdsong. The attributes of a test video are then compared against these profiles, and its location is estimated. As more videos with embedded geographical information are downloaded, the researchers will use them to train the software to recognize more and more locations.
Currently, only 3 to 5 percent of the videos uploaded to the Internet contain geographic information revealing the location where they were shot, which means that a database containing more than selected test videos will take some time to build. Even so, the nascent system is remarkably accurate. By comparing the information in the database to some 5,000 "wild, unfiltered" videos, Choi was able to pinpoint the location where 14 percent of the videos were shot to within 10 meters, or about 33 feet.