Improvements in Face Recognition

Ignore the laughable “100% accurate” claim; this is an interesting idea:

Mike Burton, Professor of Psychology at Glasgow, and lecturer Rob Jenkins say they achieved their hugely-improved results by eliminating the variable effects of age, hairstyle, expression, lighting, different camera equipment etc. This was done by producing a composite “average face” for a person, synthesised from twenty different pictures across a range of ages, lighting and so on.

Not useful when you only have one grainy photograph of your target, but interesting research nonetheless.

Posted on February 11, 2008 at 7:18 AM29 Comments


Dr. February 11, 2008 7:35 AM

I’d be worried this would lead to a lot of false positives, though… What would we get if we fed in pictures of two different people to average?

Kerub February 11, 2008 8:07 AM

actually we have a wide variety of different Bin Laden photos.

…different pictures across a range of ages, lighting…

…and so on…

Anonymous Coward February 11, 2008 9:01 AM

@ Kerub

The individuals that we have comprhensive dossiers on (Bin Laden, Zawahiri, etc.) are not the ones we should be worried about. Osama’s not going to strap on explosives and pay us a visit. Worry about the new people that they are recruiting every day. Fight their efforts to recruit Jihadi’s and stop feeding into their fear. That’s the only way to win this fight. If a young Iraqi or Palestinian feels he/she has a future to look forward to then they are much less likely to devote their life to a perverted version of Islam. Let’s address the root cause of terrorism instead of just fighting the symptoms. Lack of hope is our enemy. Why aren’t there more Iranian suicide bombers? They’re more religious than most people in the region? Maybe its because they have a functioning civil society where young people have a future.

Remember, suicide is against the Koran. They are perverting the meaning of the book to say that the only acceptable way of ending your life is to be martyred for Alah. What we’re fighting here is depression. Make is so that these young people have something to live for and the old radicals will run out of people to send to their deaths. Go Obama!

Carlo Graziani February 11, 2008 9:01 AM

This problem sounds like an interesting inversion of a standard astronomical data analysis method — comparing several images in which one feature changes (like an exploding supernova, for example), to extract information about that one feature. Here, instead, several images are compared to try to extract the stable features, while discarding stuff that changes. Kind of like swapping background and signal, in a way.

bob February 11, 2008 9:01 AM

I suspect the “100%” means that its 100% of the tests they posed succeeded.

However, I would be interested to know what the test photos looked like. Did they scan 20 of John Travolta and then use 1 of the 20 as the test? Or did they have 21 distinctly non-identical photos of him, scan in 20 and use the 21st as the test? Or did they have 100 and use 20 for training and the other 80 were tested (which would be the only meaningful result).

I wonder if at some point it is able to say “that last photo you fed in is not the same person as the previous 15” (during the training process). That would both tell you that it has adequate training and tell you you had made an error.

Anonymous Coward February 11, 2008 9:04 AM

@ everyone

Sorry for the off topic rant above. I got carried away. My nerd side says “facial recognition is pretty cool when it works.”

fh February 11, 2008 9:25 AM

I saw this article, and iit’s essentially a trivial result. Of course an average image is easier to recognize– lighting problems and any sort of temporary daily blemishes or irregularities disappear, assuming the images are at least roughly aligned with one another. If they’re not aligned, it won’t help in the first place. Everyone in my [computer vision, graphics, and robotics] lab has been wondering why the original article got accepted in the first place.

Furthermore, the average image really is unlikely to do any better in building the facial recognition system in the first place than just taking images in good lighting– and the article isn’t talking about building the system that recognizes the face, they just fed average images into a facial recognition system and found that the average images were recognized better than images that hadn’t been averaged.

Again: They did not build a facial recognition system.

Take the article with several huge grains of salt, there’s nothing useful in it.

Dscho February 11, 2008 10:28 AM

I actually work with colleagues of Mike Burton, and would like to share a few things with you, that you should keep in mind when reading this article:

  • those are psychologists. To them, a face recognition procedure involving 10 minutes manually marking up the facial features is perfectly acceptable. Sure, they have programs that do that automatically, but those programs do not work reliably.
  • this is the UK. Surveillance is all the rage, and you can forget being funded if you do not tie into that “security” hype somehow.

In that respect, “100%” gets a completely new meaning, doesn’t it?

  • these researchers are not really concerned with false positives. Recently, I personally heard a guy from Stirling’s university (which is closely connected to Glasgow, as far as face perception is concerned) talk about this incredible new system: witnesses are shown the faces of the suspects, and the facial features are varied 5% back and forth.

He was really excited, because the recognition rate was increased dramatically.

When asked, “what about the false positives?” he went quiet. Very quiet. Because there had not been any tests for that. And the system is in operation since late last year.

Go figure.

Paeniteo February 11, 2008 10:47 AM


If false positives are no problem, why don’t you suggest to them to modify their algorithm:

public String recognizeFace(Bitmap face) {
. return “Bin Laden”;

This one is guaranteed to have 100% recognition rate…

The talented Master Ridley February 11, 2008 10:53 AM

Remember, suicide is against the Koran.

I wouldn’t know because the guy who offered to get me an English Koran hasn’t delivered. Never mind – I can probably get one from the Archbishop of Canterbury.

fh February 11, 2008 11:17 AM


The problem is that there’s nothing particularly new here. It could be more useful as a result if the paper was about creating a system which took images of someone from several different camera angles in different lighting conditions (say, as a time series from several different surveillance cameras down a street), recognized the person as the same in each image based on some other cue (a reasonable tracking algorithm could do this), aligned the face in each image (not unreasonable if the face is large enough in the image, Viola-Jones is pretty reliable), then averaged and ran through a face recognition system. That would be a complete system, and a useful one.

But the averaging is the story, and average faces are not a story. Students in intro computer vision courses know about how to use average images.

Fred X. Quimby February 11, 2008 11:19 AM

All face recognition systems are unable to recognize Bruce because he is a master of cryptguise.

gatopeich February 11, 2008 1:06 PM

Smells of money-avid valiant incompetency at work.

Somebody please keep that kin away from taxpayer’s money!

dlg February 11, 2008 1:29 PM

There is a lot of interesting research going on in advanced face recognition. IMO the most promising approaches use 3D geometry to compensate for lighting and pose, for example this:

The approach Bruce found does not use 3D geometry as far as I can see, but rather eigenface fitting ( ). This technique is very powerful, but it is much less robust bad lighting and bad pose.

All of these approaches need multiple images for training, but only one for recognition, which makes them well-suited for the typical search application. Camera resolution is an issue, but I guess not for much longer.

bane February 11, 2008 1:37 PM

In addition to Dscho’s comments above, there’s another big elephant in the room in face recognition for detection (as opposed to for identity validation, ie, does this person’s face match the images on file for person P). They’re testing a database with 3628 different people. With that few people, there’ll generally be some visual characteristic that distinguishes them, so it’s not surprising you can get very high detection rates. If you have, say, 362800 people then you’re now much more likely to have several clusters of people who look very, very similar so that there’s no ROBUST visual feature that separates them. This is a fundamental problem that’s not limited to computers: the whole business of celebrity look-alikes is based on the fact that, if you have a big enough pool of people, you can find several people who basically look very like a given celebrity to human beings.

I don’t know but I’d imagine something like the airport watch list isn’t at least that big. If you have an even bigger set of people to distinguish then the problem gets worse. Again, this isn’t a computer problem per-se: police and intelligence agencies prune down the number of people that they are concentrating on detecting in any given circumstance using contextual information and expectations.

bane February 11, 2008 1:40 PM

Ugh: in my comment above it should be:

I don’t know but I’d imagine something like, say, the airport watch list is at least that big.

john February 11, 2008 1:52 PM


Standard methodology for evaluating this kind of algorithm would be to use cross-validation: divide the dataset into, say, 10 roughly random subsets, and then run the test 10 times. Each time you run the test, you pick one of the subsets to set aside as the validation set, and use the rest as the training set.

Kerub February 11, 2008 2:42 PM

mine was only a poor joke about the multiplicity of Bin Laden faces you can find in the videos.

we need facial recon that’s able to discriminate among doubles (and triples).

Mark February 11, 2008 2:56 PM

The “100%” in the abstract refers to matching 25 of 25 “averaged” male celebrity faces. (See comments above for database size, etc.)

Caleb February 11, 2008 3:01 PM

Anyone else see the huge harvesting potential sites like Facebook have in providing the necessary “twenty different pictures across a range of ages”?

Dscho February 11, 2008 5:57 PM


well, I tried to make a case against scientists without ethics. Apparently I failed.

@talented Master Ridley:

lame excuse:


I’m actually working on a viable model of the primal visual cortex (i.e. a model of your and my visual perception). A hierarchical network.

The problem is not that it will not work; it will.

The problem is: how many people I do not want to trust, not even half a second, will try to exploit the results of my research to do something that I — as an upstanding citizen — cannot at all agree with?


You haven’t got a clue on how the game is played, have you?


3D is not the magical solution, believe me. I have worked on 3D for 5 years, and tried to apply it to faces, and the results are less than promising.

Everybody who tells you otherwise just wants money to fund her research. No, really.


I could not agree with you more. But there’s something in addition to what you mentioned: being able to do such research gives you also the responsibility to use it sensibly.

For example, I have no illusions that my boss would take my research, and get a grant to work on monitoring the whole population, by all the (numerous!) CCTVs present in Great Britain.

However, coming from Germany — where we know the consequences of bound-less surveillance, and the impact on your personal well-being (not at all nice, let me tell you that; if you watched “The life of others” you will begin to understand) — I am utterly disinclined to let somebody abduct the results of my intelligence to contemptible ends.

People in the USA might begin to understand what it means when in the middle of a very personal call to your beloved one, there is an innocuous “click” on the line. And all of a sudden you cannot stop thinking that somebody is listening in on you.

And of course, you know that no decent person would take on such a job. A person taking on a job to listen in into your conversations, is a person that wants to listen in into other peoples’ conversations.

@john and Mark:

I think you got the technical side right. But that side is totally uninteresting. The implications on your life are much bigger than that.

BW February 11, 2008 9:31 PM

Reminds me of the criminal studies they did during the 19th century. They thought criminals had specific physical traits, and if they could identify them could identify criminals by just looking at them. What they did was average the photographs of multiple male criminals. The research as you can guess didn’t bear the fruit they were expecting. The resulting averages were good looking men.

moo February 11, 2008 10:55 PM

@ BW:

I remember something about a study where the participants were shown many sets of generated faces and had to select the most and least attractive faces from each pool. The characteristics of the face were carefully controlled (sizes and distances, etc… just like you can do with those face-customizing things in some modern video games). What they found was that the faces consistently rated the most attractive, had very “average” attributes (e.g. eyes not too close together, and also not too far apart). Humans seem to be wired to prefer faces that are “middle of the pack”.

passports February 12, 2008 7:07 AM

And that’s why you’re supposed to not be smiling, etc, when you get a passport photo. They’re trying to get a pre-averaged photo. Or closer to the average than it’d otherwise be.

anonymous February 14, 2008 6:02 PM

@top comment:
“actually we have a wide variety of different Bin Laden photos.”

But no moon landing photos? Ever since the fake studio moon landing tv event, I’ve been anxiously awaiting another fake tv magic bullshit event, I expect it to be something like that, some terrorist on the moon via a terrorist made shuttle or some other crap, and the public will love it, ejaculate several more billion for oil defense, i mean missile defense system.

MLK would’ve never lasted as long as he did had he been alive today, and where do you see anyone speaking directly to the people and for them to get together in peace today, like MLK, John Lennon, etc.?


They are not allowed on TV, even big mouth Rosie, when she mentioned “Gulf of Tonkin” and for people to Google it, on a popular day time tv show, was pulled from the show, which I knew she would be like anyone who dares speak to the people for the people and encourage them to think for themselves.

Which is why it doesn’t surprise me when our drinking water continue to be subjected to fluoridation, foods and drinks packed with HFCS, and drugs which should not logically be made available for the public given the rubber stamp of approval only to be pulled years later after several deaths, but the money has already been made!

No one eating mad cow meat? Other countries refusing to buy your mad cow meat? We’ve got a solution! We’ll call it the Atkins Diet, and it’ll be popular long enough to stave off the rejection of our mad cow meat! Eat up, cattle!

another day filled with more rights taken away from us, and laws being decided upon which will take even more away from us.

Eventually all of our sewage water will be monitored and information sold so they know what we are ingesting all the time and if non-violent individuals are consuming anything like marijuana so we can throw them in jail to get exposed to some diseased ridden gangland semen.

I’m still waiting for that mandatory microchip, at this point and time it wouldn’t take much for such a law to pass, and what would the sheep do? Anything, so long as it doesn’t mess with their American Idol. Perhaps the freedom chip will double as a device which eliminates the need for a tv remote control?

Nicholas Jordan February 15, 2008 10:26 AM

@from above posts:
“Let’s address the root cause of terrorism” & “fake studio moon landing tv event”

Yesterday, while taking a browse-timeout from reading “Java Cryptography Extensions” by Jason Weiss and “Algorithms in Java” by Sedgewick I happend upon a security oriented site in Canada ~ one of the links stated that a pristene / off-the-display-new box was compromised ( massively replacing system libraries ) after only twenty minutes of casual internet browsing. Another linked article cited a heavy stating ( in so many words ): “so what, why not” and claimed that breaking consumer machines was vastly more profitable today than all other previously chosen activities.

Okay, my position is that user is a twit is the equivalent of fake moon landings – but ignorance is for some reason valued in our socialization. I did a build yesterday and the usually effective public static void main(String[] args) failed to run my first attempt at cyphering. ( 56-bit DES ) I believe this is because the shell is inserting path information even in the command prompt. People will be fed what information ( papformation ) that drives consumer-based production. A nutcase html page I had up a decade ago asks: “Who controls the controllers ? Who will rule the world ?” All security breaks that I can imagine derive from someone who is not interested in how / or \ can be implemented as a TreeMap, and instead uses a trust model. No less that Jonathan Katz / Yehuda Lindell in a 2008 publication in a series ( they should know something about it ) state that Modern Cryptography ( as a formal terminology ) introduces the notion of an Honest Person. Unless you want to get a Masters in Math and attain 5+ years field experience learning how a lost key can effect a self-induced denial of service attack, consider the notion of Reality Based Cryptography: Keep your own butt out of the situation where socialization has introduced a supervisor to whom you have to kowtow. ( bowing worshipfully ) { } At the servile moment when your bills are bullying and your case is chalky, these masters of finesse value lethargic systems that are soft-realtime. So what is the ‘root cause of terrorisim’ ? This morning I wake up and the news is that an emotionless massacre was not accompanied by the band of the same name.

But the socializer who values soft-realtime insists on fake electronic realities generated for consumers and security based on a system that can be broken in twenty minutes, in other words: a belief system. Religions are a belief system, none of them have produced a producer of miracles that can be demonstrated in reality. People believe voices in a manner that is shifted when looking at a face that appears to be doing the talking. I did a moment of independent sample capturing on this about a decade ago. I had been in a hollywood video as an extra on a Country Music Television video ( Nashville does in fact value the individual integrity over commercialized backstabbing ) and had learned that HollyHeavyVideoMasters do have a deep understanding of face/voice penetration power in belief systems – religious or otherwise – to such an extent that look is a formal terminology in business / production planning. I had a beard and long hair for warmth while working in construction and it was time to take it off for warm weather. When I cut my hair down to nothing, I have an unsettling appearance that runs deep and makes the soul uncomfortable. With $800 hair and makeup, I could stimulate tv viewers. I cut my hair and beard and came in to my isp’s gaming room and sat down and started working on real cs while the others played games.

Todd ( the isp ) walked from the back toward the cash register. When he saw me, he broke. ( in the sense we use the term here ) That told me something: The power of appearance and voice is something that a belief-generation system could use to increase the power of the system from 54% to 100%, bringing the root cause of terrorism to total acceptance as a vendor of truth. Money does not drive terroristic threats and acts, it drives the system that maintains belief systems. Belief systems drive massacre, so why would the drivers put up a story that the actor is emotionless ? So that the belief systems to whom you have to kowtow can rely on emotion to market vis-a-vis having people think.

//eof – duh.

Wesley Parish February 16, 2008 3:10 AM

Some years ago, some bright spark on kuro5hin mentioned that the CIA was working on faecal recognition systems, presumably in the hope that someone would … ummm … “correct” him.

Now it seems he was right.

Hexe March 20, 2008 2:10 PM

I’ve recently twice had Canadian passport renewals returned to me because the photos did not meet the standards outlined in the passport renewal instructions. ( Once, there was a bit of shadow behind my ears, the other time had part of my face too bright.
I was a bit upset, thinking some pencil-pushing bureaucrat was being waaay too picky until I realized what was going on.
These photos need to fit within narrow guidelines for face recognition software.
But then I must ask: Why is face recognition software needed for a passport photo? Unless Canada Customs is planning on hiring visually impaired officers, there is no need for a computer to compare someone’s face with their passport photo – a human can do a far superior job to a computer.
The Canadian government must be planning on entering passport photos into a facial recognition database that will be used for surveillance. Immagine linking facial recognition with a CCT system similar to what the British already have in place. That would be a security organization’s wet dream.
There’s only one problem. I never gave my government permission to have my photo on file so I can be watched 24 hours a day. And I never will.
So – what do we do about this?

Leave a comment


Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via

Sidebar photo of Bruce Schneier by Joe MacInnis.