Forgot your password?
typodupeerror
AI

Face Recognition Algorithm Finally Outperforms Humans 68

Posted by Unknown Lamer
from the man-is-obsolete dept.
KentuckyFC (1144503) writes "Face recognition has come a long way in recent years. In ideal lighting conditions, given the same pose, facial expression etc, it easily outperforms humans. But the real world isn't like that. People grow beards, wear make up and glasses, make strange faces and so on, which makes the task of facial recognition tricky even for humans. A well-known photo database called Labelled Faces in the Wild captures much of this variation. It consists of 13,000 face images of almost 6000 public figures collected off the web. When images of the same person are paired, humans can correctly spot matches and mismatches 97.53 per cent of the time. By comparison, face recognition algorithms have never come close to this. Now a group of computer scientists have developed a new algorithm called GaussianFace that outperforms humans in this task for the first time. The algorithm normalises each face into a 150 x 120 pixel image by transforming it based on five image landmarks: the position of both eyes, the nose and the two corners of the mouth. After being trained on a wide variety of images in advance, it can then compare faces looking for similarities. It does this with an accuracy of 98.52 per cent; the first time an algorithm has beaten human-level performance in such challenging real-world conditions. You can test yourself on some of the image pairs on the other side of the link."
This discussion has been archived. No new comments can be posted.

Face Recognition Algorithm Finally Outperforms Humans

Comments Filter:
  • and/or sensationalism... a 13000 image set a whole population does not make.
    It's probably just better than the existing algorithms somewhat.

    • Re:Marketing... (Score:5, Interesting)

      by BasilBrush (643681) on Wednesday April 23, 2014 @09:50AM (#46822653)

      But Human's don't match faces based on the entire population either. Just on the faces they know.

      I don't know how many people the average person can recognise, but my guess is that it will be less than 13,000.

      This anthropologist seems to have worked in this area, and he puts the number of people you can recognise and put a name to as 1500. (You'll recognise more than that, but you won't have names to go with them.)
      http://spectrum.ieee.org/telec... [ieee.org]

    • Re:Marketing... (Score:4, Insightful)

      by shadowrat (1069614) on Wednesday April 23, 2014 @12:10PM (#46824419)

      and/or sensationalism... a 13000 image set a whole population does not make. It's probably just better than the existing algorithms somewhat.

      I doubt my internal facial recognition distinguishes between 13000 people. And worse, as damning as it sounds, it really starts to fail when dealing with people of certain races that i just didn't get much exposure to as a kid.

    • by Matheus (586080)

      That's not really the point... this is a *specific set of known hard to match faces that is used as a measuring stick for Facial Recognition algorithms. The test is checking individual 1:1 compares NOT a 1:N search. Converting a 1:1 algo to 1:N work is a whole 'nother topic. Honestly nothing that they describe in the article is new (normalizing on facial features and fixed resolution images and Vectorizing the image) so hard to see where they made their improvements. My former employer had the best Facial

  • NSA's financing?
    • by gatkinso (15975)

      NSA mission does not include IMINT or HUMINT.

      • And there are no digitized images or videos that include faces anywhere among the data they haul in?
        • And there are no digitized images or videos that include faces anywhere among the data they haul in?

          NSA hauls in no more suspects than the NRO or NOAA, which is 0. They just listen and report.

          • Listening is 'hauling in data'(unless you are purely passive about it, which they aren't.) I didn't mention humans at all.
      • I don't think conspiracy theories are needed to allege feature creep, though.

  • Shit (Score:4, Funny)

    by GameboyRMH (1153867) <gameboyrmhNO@SPAMgmail.com> on Wednesday April 23, 2014 @09:03AM (#46822221) Journal

    Of all the singularity-ish technological advances I could be alive for, it had to be this one :-(

  • by martiniturbide (1203660) on Wednesday April 23, 2014 @09:05AM (#46822233) Homepage Journal
    Let's see if it recognize me with my Groucho Marx glasses. Can authorities relate me as being a Marxist for that?
  • by Karmashock (2415832) on Wednesday April 23, 2014 @09:36AM (#46822491)

    I'm sure a computer is better at recognizing people that people don't know. That is taking in a database of 10,000 photos and matching them against a population set of 20,000,000. A human being really can't do that. However, face recognition of known faces? I really doubt the computer can beat us.

    Now, you might say the computer has an advantage in that it can look at facial statistics and match faces even after plastic surgery but that's just because those alteration technics are designed to work against human beings.

    If the primary threat is a computer recognizing you then you can alter your appearance in other ways to trick the machine. And those are technics that are unlikely to be as effective against a person.

    Furthermore, using such methods is unlikely to be as taboo in human company as doing something that fools humans.

  • As long as I can have my implanted device remind me that, yes, it's my daughter visiting today. Her husband's name is X and they're kids names are... Technology will totally save my ass someday.
    • Exactly the same technology would be very useful for salesmen.

  • by jeffb (2.718) (1189693) on Wednesday April 23, 2014 @09:50AM (#46822667)

    I was hoping the link would include an actual self-administered test, with your score reported at the end.

    I looked through the "same" pairs a bit. It confirmed that I'm terrible with unfamiliar faces -- on perhaps a third to a half of the pairs, I would've had no idea that both represented the same person. For faces that I've seen hundreds or thousands of times -- Condi Rice, Rumsfeld, Bono, Jimmy Carter -- I guess I do as well as anyone.

    For the "different" pairs, I only looked through the first dozen or so, and they all seemed obviously different to me. Also not surprising; I get lots of false negatives (failing to recognize a face), but few false positives (thinking I recognized someone when I really didn't).

    I wonder how my life would've turned out differently if I'd grown up with a prosthesis to help me recognize faces. I wonder how much difference it would make if I got one today. I feel like I've come up with reasonable coping strategies, but I wonder...

  • Now where can we get it... Or purchase it.
    Is it patented, is it Open Source.
    Where is the link to the actual algorithm.

    • Now where can we get it... Or purchase it.
      Is it patented, is it Open Source.
      Where is the link to the actual algorithm.

      The second link is to the article. It's available as a PDF. Only use it for good.

  • by koan (80826)

    Complete and utter oppression of humans moves one step closer, soon no escape ever.

  • It's really not that hard to beat me : I say "Hello, my name is ..." twice to almost everyone at social events.
    Bogosorting pictures would do a better job than me.

  • If you have a 97.54% chance of a "perfect" match on all those criteria, remember the birthday paradox. If you compare 23 people against one another, you get (23*22)/2 = 253 comparisons. Multiply 253 * 0.9754 and you get 246.7 correct, 6 wrong.

    Six failures out of a pool of 23 (hijackers + passengers) is insane, and is therefore one of the reasons that the German Federal Security Service rejected my employers' facial recognition system many moons ago. Until we get to 1.0, the number of false positives wil

    • If you're looking for terrorists, it much worse than what you say.

      How many terrorists are there really?
      For arguments sake I'm going to say 10,000 in USA, with 320M population.
      To get the number of false positives on the general population down to the number of real terrorists, so only 50% of the people you stop are innocent, you need a 99.996875% accurate test.

      With 98.52 as claimed by this test, you'd have 4.7M false positives and 148 terrorists get false negatives.

  • Either facial recognition software has gotten a whole lot better, or the legalization of marijuana has started having an effect on people.

    I'm willing to believe both stories.

  • by Anonymous Coward
    Such a good result would easily get published on a high impact journal or conference. If they are self-publishing on arxiv, there is something very fishy about the research. No proper review, no source code, and chinese authors (known for flooding journals with incorrect bullshit in hope of gettings something publushed by accident).
  • Now the next big challange for AI will be channel surfing. I doubt it will ever change channels as fast as men do and still know what is being shown on each channel.
  • by HannethCom (585323) on Wednesday April 23, 2014 @11:22AM (#46823801)
    Both the title and description are sensationalism. If you read the paper, the algorithm works at 98.52 on 1 of 5 data sets. We could also accurately say the algorithm is only 89.33% accurate. The score from the most difficult database. Much worse than the normal person.

    That being said, it is much better than other publicly available algorithms.
  • by 140Mandak262Jamuna (970587) on Wednesday April 23, 2014 @11:35AM (#46823973) Journal
    One of the most prominent features of Lord Shiva, is His third eye. [google.com] Paint a third eye on the forehead to completely discombobulate Gaussian face.

    The other stunning form of Lord Shiva is His half-female version [wordpress.com]. If you could manage this form, you would discombobulate not just Gaussian Face but also fellow humans too.

    Extending the theory, painting random noses, lips, eyes and other features on cheeks, foreheads etc would defeat these automatic face recognition systems.

  • It's a lot of rather common techniques put together, but it's a meaningful result. But the summary is rubbish: on the same benchmark for which this algorithm gave 98.52%, there was an older algorithm that gave 96.33%, and a graph showing small incremental improvements up till then; saying that older results "never came close" is a blatant lie.

    The methodology seems legit, in the sense that it's consistent with those used for previously accepted results. On the other hand, I still have the same old concern o

  • The machine margin is probably fairly exact, but measuring humans gets a bit fuzzy. What is the margin of error on that 97.53 percent for humans. If it's 3% then I'm going to doubt the claim.
  • of Killbots thank you.

    "Science: we're all about coulda, not about shoulda." -- Patton Oswalt
  • Imagine a tech savvy guy planning to commit some serious crime. He/She could find a patsy, take a few photographs on the sly. And get to work at home subtly adding makeup to make the eye slightly apart, or contact lenses with off-center pupils, add lipstick to make the lips slightly longer etc. Keep altering it tell Gaussian Face declares that new face of the criminal and the photo of the patsy both hash to the same 150 pixel canonical image. Now, commit the crime in full view of the security camera, and th
    • Except I doubt 98% accuracy is admissible in court.

      Say there are 320,000,000 people in a country, like USA.
      If 5% of those people are criminals, that's 16,000,000 criminals.

      With 98% accuracy against the 304,000,000 innocent civilians, that's 6,080,000 false positives, nearly half of the number of real criminals. That's only 62% change of successfully identifying a real criminal, despite a "98% accurate" test.

  • So to fool the face recognition of the future I just have to change the distance between the corners of my mouth.... by smiling or frowning?
    If I perform all my crimes with a cheeky grin, they'll never match it to my frowny looking drivers license photo.

  • It might not be as good as a human but quantity has a quality of it's own. If cameras can pick out likely suspects then a human can take a closer look from the same feed.
  • Ah, so would it recognize that Clark Kent and Superman are actually the same person? Imagine the benefits for villains all over the world!

Almost anything derogatory you could say about today's software design would be accurate. -- K.E. Iverson

Working...