Tuesday, February 13, 2007

Face Recognition for Online Photo Searches Sparks Privacy Fears

A new type of search engine using facial recognition technology could soon be able to pinpoint images of a person among the billions of photos posted online—even if their name does not appear.
A Swedish company named Polar Rose plans to launch its service for facial searches tied to the photo-sharing site Flickr within a couple weeks.

In the next few months the firm hopes to expand the service to search images across the entire Web.
The technology promises enhanced photo finding that would make it easier to find people on the Internet.
But privacy advocates are concerned that Polar Rose and similar facial-mapping search engines will violate people's rights and potentially aid criminals.
Lee Tien is an attorney at the Electronic Frontier Foundation, an Internet watchdog group that focuses on privacy and civil liberties.
"Photos [posted online] are effectively anonymous now," Tien said, unless they are labeled with some sort if identifying text. "But if Polar Rose works the way they say it will, that's all going to change."
Tien said that this kind of technology could aid stalkers in tracking down their victims, or it could allow employers, insurance companies, and the government to pry into people's lives more than some of us would like.
(Related news: "Privacy Fears Intensified by Tech That Knows Where You Are" [October 20, 2006].)
"People are going to have to be more careful about picture-posting and figure out ways they can prevent pictures from being [analyzed] by these search engines," Tien said.
User Input
Current image searches are based on the text surrounding a photo. Images of your neighbor, for example, should only be found if his or her name is in an associated caption or article.

Polar Rose's system works by analyzing a photo to create a three-dimensional sculpture of a person's face, which can account for changes in angle and lighting.
The company developed a computer system that takes about a second and a half to roughly guess what a person looks like in 3-D based on a photo.

The service then encodes the 3-D portrait as a set of numbers describing the face, which they call a "face print." Different face prints can be compared to see if they appear to be the same person.
The system is more than 90 percent accurate when combing through a set of about 10,000 photos, Nyholm said.
But the Web contains vastly more photos, so there are more images of people that look similar enough to trick the system.
"If we searched purely on face prints, it wouldn't be reliable enough," said Polar Rose CEO Nikolaj Nyholm. "That's why we need the help of users."
Polar Rose's system will ask users to input the names of people they recognize in photos online.
With users' help, the company will build a huge database linking together various photos of a person along with other information about them.
This database will then be able to sift through the billions of online photos and tell the difference between, for example, a girl and her sister.
"We do believe we'll be harnessing the collective intelligence rather than just the intelligence of single users," Nyholm said.
Nyholm added that Polar Rose is concerned about privacy.
The firm plans to give people ways of removing photos from the company's database or to avoid having their online photos analyzed and catalogued in the first place.
Reliable Results
Privacy fears aren't the only potential obstacle for face-mapping searches.
Reliable search results are also an issue, and it remains to be seen whether Polar Rose's service will work as well as advertised.
Riya, another company experimenting with visual searches, abandoned their face-recognition project, in part because it wasn't working well enough to meet people's needs.
"A million years of biology have trained you to be a face recognition machine," said Munjal Shah, CEO of Riya.
"You are better at this task than almost any other task you do, and I'm not sure computers are ever going to meet the satisfaction of users."
(See results from a 2005 survey testing how well people can read emotions in the faces of strangers.)
Instead the firm offers a service called Like.com, an image-searching tool that sifts through online catalogs of things such as shoes, handbags, and jewelry to look for items that have similar color, shape, texture, and other qualities.
The site can even pinpoint items in photos—such as a pair of designer shoes that starlet Lindsay Lohan wore to a movie premiere—and search online stores for look-a-like products.
"For these aesthetic items, they're very hard to describe in words," Shah said, and different people use different words for them.
This means, Shah contends, that these kinds of things—not faces—are best suited for visual search technology.