A Swedish company will launch an Internet search technology next month that uses 3D modelling to find photos of people based on their facial features.
Millions of photos are added to Web sites such as Flickr.com and others daily, but only the text metadata is searchable, which is often misleading, said Nikolaj Nyholm, chief executive officer of Polar Rose.
The Malmo, Sweden, company, founded in November 2004, was working with Western European governments who were interested in the technology for law enforcement. But Nikolaj said the company decided in March to end that work and make the technology widely available.
Polar Rose uses software spiders to mine the Internet for photos of faces as well as accompanying text. When a face is found, 90 different measurements are taken, based on 3D data and texture characteristics, Nikolaj said. A 90-character alphanumeric string is then created. Faces with as few as 100 pixels can be recognized.
Polar Rose uses Amazon.com's Elastic Compute Cloud (EC2) data-processing service, Nikolaj said. The data is then sent to Polar Rose's databases in Denmark and Sweden. The company is currently building up a larger database of photos across the Internet.
Polar Rose is also creating a plug-in for the Firefox and Internet Explorer Web browsers. The plug-in, which the company expects to release by February, pops up a small symbol when the browser opens a Web page containing a photo of a face. When a user clicks on the symbol, the Polar Rose service looks for matching photos in its database, Nikolaj said. Users also can tag existing photos in the database to help train Polar Rose's matching engine.
Photos marked as private -- for example, in photo-sharing services -- can't be searched. The accuracy of the application is between 92 percent to 95 percent in controlled environments with image sets of no larger than 10,000, Nikolaj said.
The accuracy declines with larger sets because people start to look alike, he said. But then the search engine uses metadata associated with that person's face to come up with a ranking of selections, which are then displayed in a browser window, Nikolaj said.
"We really see this as harnessing our users' collective intelligence," Nikolaj said.
Polar Rose will license its APIs (application programming interfaces) royalty-free to its partners, Nikolaj said. Its first partner, a European photo-sharing site, will be announced in January, he said.
"Flickr is probably a good model for who we want to reach," Nikolaj said. "We'd love to work with them, obviously."
The plug-in will work independently of whether a Web site has implemented the Polar Rose technology. But Polar Rose sees partner arrangements as a way to eventually incorporate contextual ads, Nikolaj said.
For example, if a person is searching for a photo of Bono from the band U2, links to an MP3 seller could appear, he said.
In mid-January, Polar Rose will release a preview program, called Limelight, that uses the same technology as its forthcoming plug-in. After a person has uploaded two photos of themselves to Polar Rose's database and supplied their Flickr ID, Limelight will find photos of that person in the person's Flickr friends.
Limelight can then e-mail the results or supply them via an RSS (Really Simple Syndication) feed, Nikolaj said.