Print

Mail

wiki links

A faceprint in the crowd

By Ronan Fitzgerald
THE GUARDIAN, LONDON
Monday, Jan 29, 2007, Page 9

Jim Greer is a games programmer who wanted a pair of vintage running shoes. He knew what he wanted, except he wasn't sure where he'd find them. He ended up using an image search engine called Like.com to find something similar.

"I searched for the model I wanted and was able to single out the shape of the toe as the feature I liked the best, and find more similar shoes."

Cue one satisfied customer. But what's strange about it? Because although being able to search for an object that "looks like" another feels natural to us, computers can't do the same.

They can compare strings of bits very quickly -- and so compare two files of text in an instant. But rotate an image by 90 degrees, and the stream of bits comprising the image file changes too. Yet the object's the same.

Take faces. Humans can tell faces apart very easily, or recall that they've seen one. Computers can't. But that could change with the launch of a new image search engine developed by the Swedish company Polar Rose (polarrose.com).

This works via a plug-in for the browsers Firefox and Internet Explorer, which combines computer analysis and social tagging; if it succeeds, once your computer has seen one photo of you, it will be able to recognize you again in countless more pictures.

Furthermore, any Polar Rose user who takes a shine to one of your photos -- that self-portrait on Flickr, perhaps? -- will be able to search the web for "more like this," at the slight risk of finding your twin or just someone with an uncanny resemblance.

Poor track record

Mikkel Thagaard, vice-president of business development at Polar Rose, explains: "Anywhere you go online [using the plug-in], what you'll see are rose icons popping up wherever there's a face. When you click on these roses, our technology makes what we call a faceprint.

You're familiar with a fingerprint. This is the same thing, but for a face.

Then the user can look elsewhere on the web to find matches for that particular face." Polar Rose will also let people tag faces with names, which he hopes will improve its accuracy.

Thagaard believes that traditional image searches, which actually search for text tagged to the image or nearby, are too inaccurate on an increasingly visual web.

"There's no way to search for images unless they're tagged, and even then you can't be sure you'll find the image you want because there are two people's interpretations at work. The person tagging the image has to have the same idea of the image as the one that is searching". The difference with Polar Rose or other image search engines is that you are searching for images by pointing to an image you have already and saying, "find me more like this."

Since Polar Rose is still only in beta, we can't be sure that it will work. Indeed, facial recognition technology has a poor track record.

In 2002, Florida police withdrew crowd surveillance technology from the American company Visionics after just two months -- it had failed to identify a single criminal on the police database while raising a number of false alarms.

Polar Rose is going out on a limb by claiming to have a system that works. But even Thagaard admits that facial recognition technology is still learning.

"Looking at the photos on the web, there's bad quality, different expressions, different poses and different light sources, so it's not easy. Even for people it's hard to distinguish between two photos sometimes."

Google

Furthermore, it may soon have some competition from Google, which last year acquired Neven Vision, a company with a string of patents in the image recognition field. It seems likely that the dominant search engine will want to offer more accurate image searching.

Image recognition could make online shopping easier too. That's certainly Like's aim, where you can search for products based on their color or design. If you've broken a plate or can't find your favorite shirt, simply take a photo, upload it and Like's search engine will try to find matching items in online stores.

You can highlight a part of an item that most appeals to you and even specify whether you want the search results to concentrate on items of the same color, or items of the same shape.

Munjal Shah, chief executive of Like, explains how it works: "We convert each picture into a math formula, and we call this maths formula a visual signature. The visual signature is about 10,000 numbers from each picture, which encodes in it the shape, colour and texture of an item, and all the other details.

"When you do a search what you're actually doing is comparing the visual signatures of two different items. So strictly speaking, we don't compare the two pictures, we compare the two maths representations. It's all done using these mathematic visual signatures," he said.

There's not much to fear about image recognition when it's helping us find a polka-dot tie. But when the focus is on faces, things aren't so rosy.

Granted, a search engine that recognizes faces could make managing images and searching for them a lot easier. Instead of tagging photos, you could simply use one photo of yourself, or whoever, to find all the others. But the problem is, so can anybody else. Just as it didn't take long for people to become worried about "being Googled," if image search engines take off we might have to be a lot more cautious about the photos we share online. By syncing photos of us with whatever other personal information we've left lying around online, who knows just how detailed a personal profile could become?

Surveillance society

Jonathan Bamford of the UK's Information Commissioner's Office, the independent body that both promotes access to official information and campaigns to protect personal information, believes this is a genuine concern: "New technology means the risk of constant surveillance is higher, and we need to be aware of that risk. We are concerned that we're putting together the infrastructure for a surveillance society."

However, Shah, who is also chief executive of Riya (riya.com), the first photo sharing site to use facial recognition technology, thinks faces are still beyond computers' reach. "We tried the same algorithims that are used on Like.com for faces on Riya. We downloaded 50 million faces from MySpace, and we brought in 100 users to test both the shopping and the faces. The face stuff got 2.5 out of 5 on average, the shopping stuff got 4.2 and up," he said.

"There's something about faces. You can't win. It isn't about the technology. The technology is performing at the same level of efficacy as it is with items, it's just our expectation is substantially higher because something in our brain is better at it," Shah said.

Thagaard doesn't see the potential for abuse, either: "We're trying to give people a solution that will help them, and we have no benefit from making something that people will abuse. We will always be very aware of privacy concerns people have and be open to hearing them."

Might he be right? Time and again we see that where technology provides neat time-saving applications, people are willing to give up personal information. But Bamford believes this must change. "I don't think people understand that the more and more you put out there in the electronic world, the more people can potentially access that," he said.
This story has been viewed 1314 times.