Eigenfaces Online Service

nordicfrost writes "At Norways largest online newspaper, they're doing the beta test of an Eigenface service. In short, it takes a portrait picture, treats it with some filters and analyzes the vectors in it. This is used for crimestopping and generic Big Brother activities. In this database, however, your mug is compared with celebrites to find which one that matches you the most. I match, among others, Brad Garrett. This site is pretty self-explaining, upload a portrait pic of yourself taken right in front of your face, closely cropped to the face and use the up arrow to see your top ten list of celebrity matches. You have to agree to some terms first, like giving permission to use your pic in promotions and other stuff and not to upload indecent stuff."

Comments

William Gibson novel

[raider_red]

Wasn't there something like this in Virtual Light? I think it was used to find missing persons by posting information on which celebrities the abducted/missing person looked like.

Avoiding recognition: Gaming the system

[G4from128k]

A facial recognition system like EigenFace looks for the dominant modes of variation among a population of faces to find predictors that help distinguish among the different faces. By ignore the ways that faces are more similar and accentuating the ways that faces are most different, they can craft identification algorithms. The only challenge is getting enough data to understand the full range of facial variations -- hence this solicitation for more facial data.

But I wonder if people can game this system to make themselves unrecognizable. For example, a member of the tin-foil hat brigade might submit multiple mugshots of themselves under mutliple assumed identities. By using slight variations in facial expression, makeup, lighting, and camera angle they would make the system think that a large fraction of the population "looks like" the person who seeks anonymity. The system would then have a hard time identifying the tin-foil hat wearer because he/she matches so many people in the test data.

Anyone can make themselves "look average" if they can bias the dataset that defines the average.

If only..

[Willeh]

I just hope they can keep this out of the hands of the police. They would LOVE to get their hands on a database like this (even with the inevitable false positives you get when people just want to prank the system with mcbride/ hello.jpg type stuff). If you think i'm being too serious about this, think of the blurred lines between the media and serious business that is already the case in some countries. Keeping in the vein of the article, spot the quote: "Now i can feel like Matt Dillon instead of an asshole."

Medical application

[InternationalCow]

Perhaps less obvious is the usefulness of this system for syndrome diagnosis. There, we depend on our visual memory of a dysmorphic "Gestalt" to recognize human malformation syndromes. So far, attempts at some kind of computer-aided diagnosis have failed. Too bad, because syndrome diagnosis like we do it now is error-prone and difficult. This system however may be used to recognize typical "syndrome vectors", so you can feed it a patient's picture and have it say "x% match with oculo-dento-digital dysplasia" or whatever disease comes to mind. Potentially very useful!

Re:Medical application

[Anonymous Coward]

Absolutely fascinating idea.

I've worked in neuropsychology and neurology, and there's this phenomenon that often occurs where you meet a patient, and know they're dysmorphic in an identifiable way, but can't quite place what's different about them. These individuals often go through years, if not decades, of unsuccessful genetic testing to try to match them with a syndrome.

Perhaps your suggestion could help in such cases. More generally, it might improve research, by allowing for quantification of features more accurately. Then you could say with some precision that feature X tends to be associated with gene Y. This might actually lead to a revolution in the way that neurogenetic syndromes are conceived of and neurodevelopmental trajectories are modeled.

Having said that, I'm not sure that an eigen decomposition approach is the most effective. As I understand it, this method is based on a decomposition of a covariance matrix, and is therefore essentially limited to breaking down features into linear components. It only models through the second moment of a distribution (means and variances).

My guess is that more accurate results could be obtained by using something that models statistical dependencies in general, rather than just covariances. This could be done by using independent components analysis, for example, rather than principle components analysis.

Re:Ima Gunna (Test it)!

[gosand]

Better yet, test the site out by uploading pictures of celebrities, and see which celebrity is says they most closely resemble.

Let's be honest for a second...

[kill-9-0]

How many of you REALLY REALLY want to upload a picture of the goatse guy and see who he most resembles? Any guesses?

My use for this

[FuzzyDaddy]

I could use some decent facial recognition software to help me sort my rapidly growing collection of digital family photos. I barely have time to pick out the good ones for printing, let alone organize them in some rational way. I've often wished I could use facial recognition software to go through and create a database of the people in my pictures, so I could automatically send pictures to people who would be interested.

An extended use

[fireacc]

It would be interesting to have them not limit the matches returned to just celebrities, but to return matches to other users of the system. Then you could see if you have an evil (or good) twin out there somewhere.

Someone's editing wikepedia again...,

[Eric(b0mb)Dennis]

Eigenface

From Wikipedia, the free encyclopedia.

Eigenfaces are eigenvectors in the high-dimensional vector space of possible faces of human beings. These eigenvectors are used in the computer vision problem of human face recognition.

Some eigenfaces.
Eigenfaces also are linked to the slashdotting of many sites.

jaj

Egads! Scary stuff....

[dinohum]

Tried it with a picture of Adolf Hitler, said it matched David Hasselhoff....then tried it with Margaret Thatcher and it matched Naomi Campbell...surely there is somebody with a sick sense of humor really making the matches. Though then again, I do hear Hasselhoff is popular in Europe.