The Scandinavian company's recognition technology could make Web image searches a lot more effective
A picture may be worth a thousand words, but when you use words to search for pictures on the Internet, the result is often worthless. Try looking on Google Images for, say, photos of your kids. Entering my daughter's name returned a group shot of some hockey players and a snap of a burly bear. Querying the name of a female colleague yielded images of Dick Cheney, a UFO, and a fat Buddha. (Fortunately, my office mate had the good humor to find it hilarious.)
Scandinavian startup Polar Rose, named a World Economic Forum Technology Pioneer on Nov. 29, aims to improve on such feeble and often amusing results by employing the latest technology in image recognition to help users find pictures based on what they depict, not just on the titles or captions attached to them. It's an ambitious goal—something of a Holy Grail in computer science—but Polar Rose already has achieved breakthroughs worthy of recognition.
The Web is becoming increasingly visual: Photos and video are the fastest growing type of online content. About 10% of Google (GOOG) searches now are for images, says Nikolaj Nyholm, the 32-year-old Danish chief executive of Polar Rose. Yet, he adds: "There are very few innovations around improving that search, so there is a tremendous opportunity here for Polar Rose."
Existing search services such as Google, Yahoo! (YHOO), and MSN (MSFT) locate images by scanning the text attached to photos. The problem, of course, is that the accuracy of the search relies on how well images are tagged, if they're tagged at all. Polar Rose has devised technology that scans the images themselves, sensing the shapes and comparing them with a database of known images. It's being used primarily for facial recognition but later will be expanded to recognize animals, cars, and other objects.
To be sure, Google and others aren't standing still. Last year, Google bought Neven Vision, a photo-recognition firm in Santa Monica, Calif., with a background in biometrics. And Microsoft has acquired a startup called SeaDragon Software and is collaborating with the University of Washington on a project called Photosynth that aims to analyze digital photos for similarities. But so far, Polar Rose looks to be ahead.
The company grew out of work at Sweden's Malmo and Lund universities, where researcher Jan Erik Solem was developing surveillance technology for monitoring people in train stations. The difficulty of recognizing faces in photos is made all the more difficult by changes in lighting and pose, both of which can drastically alter a person's appearance. Solem's breakthrough was technology that extracts a virtual 3D portrait from a single two-dimensional photo. Then, even if the subject turns his head, the 3D representation is accurate enough to recognize him.
The idea of using Solem's technology to spot faces on the Internet came from Nyholm, a lifelong techie who has lived in India, Africa, Europe, the U.S., and Costa Rica. He was casting about for a new project when a venture capitalist asked him to advise Polar Rose. Soon after, he came on board as CEO and pushed the company to turn its talents to online image recognition.
One key to how Polar Rose works is that it taps into the so-called collective intelligence of the Web. Users of the product, which is still in beta testing, help improve it every day by confirming or correcting its search results. Thus, if they see a photo that is clearly Jeane Kirkpatrick but tagged as Alexander Haig or Bobby Inman, they can make Polar Rose "smarter" by fixing the mistake.
Users also constantly add to Polar Rose's database by submitting their own correctly identified photos, which Polar Rose can then compare with millions of others. Upload your wedding pictures with all the guests identified by name, and suddenly you can find other snaps of the same people elsewhere on the Net. Nyholm says Polar Rose's 1,800 beta testers are already viewing and verifying a half million photos a day. To his amazement, viewers spot people they know in about 3% of the pictures they see, so the database of confirmed identities is growing fast.
Nyholm expects to launch the service publicly in 2008. In the first quarter, Flickr users will be able to use Polar Rose to automatically tag photos in their personal albums. And around the same time, Polar Rose will introduce an interface that lets any Web site—from individual blogs to large social networking sites—offer the same capability by adding a few lines of computer code to each page. If Polar Rose delivers as promised, looking for images online should get a lot easier—though perhaps a bit less funny.