Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 



Forgot your password?
typodupeerror
Compare cell phone plans using Wirefly's innovative plan comparison tool ×
AI Input Devices Software Yahoo! Technology Science

Breakthrough In Face Recognition Software 142

An anonymous reader writes: Face recognition software underwent a revolution in 2001 with the creation of the Viola-Jones algorithm. Now, the field looks set to dramatically improve once again: computer scientists from Stanford and Yahoo Labs have published a new, simple approach that can find faces turned at an angle and those that are partially blocked by something else. The researchers "capitalize on the advances made in recent years on a type of machine learning known as a deep convolutional neural network. The idea is to train a many-layered neural network using a vast database of annotated examples, in this case pictures of faces from many angles. To that end, Farfade and co created a database of 200,000 images that included faces at various angles and orientations and a further 20 million images without faces. They then trained their neural net in batches of 128 images over 50,000 iterations. ... What's more, their algorithm is significantly better at spotting faces when upside down, something other approaches haven't perfected."
This discussion has been archived. No new comments can be posted.

Breakthrough In Face Recognition Software

Comments Filter:
  • Upside Down? (Score:5, Insightful)

    by Anonymous Coward on Tuesday February 17, 2015 @09:10PM (#49077381)

    "What's more, their algorithm is significantly better at spotting faces when upside down, something other approaches haven't perfected."

    Add this step: Rotate the image and run the algorithm each x degrees. What am I missing?

    • Re:Upside Down? (Score:5, Insightful)

      by kekx ( 2828765 ) on Tuesday February 17, 2015 @09:45PM (#49077611)
      Performance.
      • a worse case scenario of x4 times worse performance...and/or battery life if working on a mobile.

      • do it in parallel in hardware. There are FPGA and ASIC solutions that can do hardware rotation, just send one through the rotate matrix into identical hardware. It costs twice as much, but in todays terms that still shouldn't be too bad.

    • False positives
    • We are finally going to catch this guy!! - http://img.izismile.com//img/i... [izismile.com]

      (the problem is the background - your brain is very good at understanding what upside-down means, but an algorithm trained by seeing tons of right-sided up images only understands that a silo is rounded on top and straight on the bottom. - The question I have, is what are the pratical implications of all the extra processing power that might take? Finally figuring out who that gymnast was from that circ-du-soleil screenshot? )

    • by mcrbids ( 148650 )

      There's lots that you are missing.

      The issue isn't the input data, it's the processing method. The processing method mentioned here as "revolutionary" is just about exactly the method that Raymond Kurzweil [wikipedia.org] posited: a hierarchy of "nodules" that pattern match on a cascading network of pattern matches....

      We're living with a modern-day Turing. Do we give him ample credit?

    • by Anonymous Coward

      Finally! We can bring the benefits of surveillance to southern hemisphere countries...

    • by Anonymous Coward

      The faintest grasp on machine vision. That goes for your 5 moderators, too.

    • by duhjim ( 733407 )

      "What's more, their algorithm is significantly better at spotting faces when upside down, something other approaches haven't perfected."

      Add this step: Rotate the image and run the algorithm each x degrees. What am I missing?

      That this is also true for all of us humans when we exit the womb?

  • "Breakthrough in face recognition software"
    "The idea is to train a many-layered neural network using a vast database of annotated examples"

    How novel.
    • by Tablizer ( 95088 )

      It begs the question: why were they using few layers and skipping annotation in the past? The hardware couldn't handle it? They were too lazy to implement such? They needed a Flux Capacitor to make them work together? The boss didn't like the "look and feel" of the diagrams? It crashed Windows XP?

      • The annotations were probably more useful features such as distance to the subject, angle of head tilt, or principal lighting angle, lens focal length. Train a net to recognize the shape of heads tilted at various angles and you've gone a long way toward recognizing faces tilted at those angles. Now you can train separate networks to recognize faces at each specific angle or small range of angles. The same for dealing with varied distances and lens focal lengths.
      • by ceoyoyo ( 59147 )

        There wasn't a good algorithm for training general deep ANNs until 2006, although convolutional neural networks were an exception to that. It's likely nobody tried it before because computers weren't fast enough and the discovery of layer-wise unsupervised training hadn't made deep networks popular yet.

      • Re:so breakthrough (Score:5, Informative)

        by tmosley ( 996283 ) on Tuesday February 17, 2015 @10:34PM (#49077831)
        It seems to me, as I have been following the progress of the technology over the last year or so, that it was only recently that scientists either had the idea to layer networks on top of one another, or gained the ability to. This started with the algo that would analyze pictures for content and tag them, ie a picture of a girl playing with a dog was tagged as such. It was approaching primate-level "cognition" in that specific context a few months ago, but now I have read that it has reached or surpassed peak human level, where rather than labeling the dog as a dog, it labeled it as its specific breed, or labeled a flower as its specific type that I had never heard of. Combining that with this new data point, it would seem that visual perception in machines has exploded into post-human territory. Shit is getting real.
        • by Anonymous Coward

          Attention technologists: quit inventing stuff that enables corporations and governments to spy on people!!!

      • Re:so breakthrough (Score:5, Interesting)

        by hughperkins ( 705005 ) on Tuesday February 17, 2015 @11:16PM (#49078041) Homepage

        They're using a standard technique. Convolutional networks started to become big with LeCun's 1998 paper on learning to recognize hand-written digits http://yann.lecun.com/exdb/pub... [lecun.com] . His lenet-5 network could identify the digit accurately 99% of the time.

        Convolutional networks are starting to become used to play Go, eg 'Move evaluation in Go using Deep Convolutional Neural Networks', by Maddison Huang, Sutskever and Silver, http://arxiv.org/pdf/1412.6564... [arxiv.org] Maddison et al used a 12-layer convolutional network to predict where an expect would move next with 50% accuracy :-)

        Progress on convolutional networks moves forward all the time, in an incremental way. If we had one article per day about one increment it would quickly lose mass appeal though :-) The article is about one increment along the way, but does symbolize the massive progress that is being made.

        Convolutional networks work well partly because they can take advantage of the massive computional capacity made available in GPU hardware.

  • by Snotnose ( 212196 ) on Tuesday February 17, 2015 @09:14PM (#49077401)
    For every "terrorist" they track through the mall, how many ordinary Joes like me who like their privacy are also tracked and stored in huge databases for all time?
    • Yeah, I was surprised there was no mention of the huge privacy implications this has. But hey, maybe this'll reduce the number of IDs and RFID cards you have to carry around since it'll be so easy to identify and track you when you're just walking around.

    • by RoknrolZombie ( 2504888 ) on Tuesday February 17, 2015 @09:24PM (#49077481) Homepage

      All of them.

      • by retroworks ( 652802 ) on Tuesday February 17, 2015 @10:00PM (#49077679) Homepage Journal
        "For every "terrorist" they track through the mall, how many ordinary Joes like me who like their privacy are also tracked and stored in huge databases for all time?"

        Indeed, all of them.

        Have you noticed you can go into Best Buy or Staples, pick up a camera or look at a printer you never searched for online, and you find ads for the device on Facebook? Didn't notice? Give it a try. It's far beyond this 2013 (minority) report http://www.businessinsider.com... [businessinsider.com]

        • No, for that I'd have to go into a Best Buy or Staples, and then use Facebook.

          I kid, a bit, because I have actually been into a Staples recently, but since they were nowhere near having what I wanted or prices I would pay, I don't think I'll repeat that. I just needed one final reminder that it's a waste of my time.

    • how many ordinary Joes like me

      Why, all of them, of course.

      And should you ever commit a crime we will be able to retroactively find the evidence for your trial. If you really piss us off we'll edit the video record and call it parallel reconstruction.

      In a few years, the pre-cog program will come online, but the surveillance is here to stay.

      Now stop picking your nose, citizen.

    • It's recognition, not identification.
      As in a yes/no if an image contains a face. No who is in the image.

    • by Jeremi ( 14640 )

      It's kind of beside the point whether it's a good thing or a bad thing. No doubt it will have some combination of good and bad effects, but regardless of what the effects are, the cat is out of the bag -- the algorithm is invented and it's not going to go away. And if these guys hadn't invented it, somebody else would have. The only question that remains is how society ought to react to its existence.

      • The question is whether we should allow government to scale to be big enough for it to be a powerful tool.

        We can clip some wings by not allowing ubiquitous cameras, or by limiting how big powerful global organizations can use the tech.

        It isn't inevitable due to the existence of the technology. The technology exists for mass low cost execution of people. We don't allow large overreaching organizations to execute people freely. It remains a rarely used technique. Restrictions on the scaling of face recogni

    • Indeed. I am so torn over this. On the one hand, the technology is very cool. On the other hand, the inevitability of abuse seems to outweigh the benefits.

  • Spike boots (Score:5, Funny)

    by Tablizer ( 95088 ) on Tuesday February 17, 2015 @09:15PM (#49077411) Journal

    What's more, their algorithm is significantly better at spotting faces when upside down

    Rats, there goes my ceiling-walking bank-robbery plans.

    • If you wore a mask that made your face not look like a face, it will ignore you.
      • Or a mask with someone else's face on it, Or a T-shirt with a few faces on it, or a baseball cap, or a burqha...
        • Re:Spike boots (Score:4, Informative)

          by hughperkins ( 705005 ) on Tuesday February 17, 2015 @11:05PM (#49077989) Homepage

          Yes, check this out 'High Confidence Predictions for Unrecognizable Images', by Nguyen, Yosinkski and Clune, http://arxiv.org/abs/1412.1897 [arxiv.org] . It's a paper that shows an image that the net is 99.99% sure is an electric guitar, but looks nothing like :-)

          For the technically minded, the paper's authors propose that the reason is that the network is using a discriminative model, rather than a generative model. That means that the network learns a mathematical boundary that separates the images that it sees, in some kind of high-dimensional transformed space. It doesn't learn how to generate such new images, ie, you cant ask it 'draw me an electric guitar' :-) Maybe in a few years :-)

          The authors don't compare the network too much with the human brain though, ie, are they saying that the human brain is using a generative model? Is that why the human brain doesn't see a white noise picture, and claim it's a horse?

          • by ceoyoyo ( 59147 )

            There are two popular types of deep ANN at the moment: restricted Boltzmann machines and auto-encoders. RBMs are generative. Autoencoders can also be generative if you train them in a particular way, which works much better so most people train them that way anyway. So you can take an ANN and ask it to draw you a picture of a guitar.

            I disagree with the authors of that paper. It seems more likely to me that they've cherry picked particular examples that fool their particular ANN. That's pretty easy to d

          • The authors don't compare the network too much with the human brain though, ie, are they saying that the human brain is using a generative model?

            I don't think so, because saying something like that is not supported by evidence. The human brain doesn't actually work like neural networks do. Neural networks are only loosely inspired by one very, very narrow and specific aspect of the mechanics of the brain.

            Is that why the human brain doesn't see a white noise picture, and claim it's a horse?

            The human brain does this sort of thing all the time. You can see shapes in static, of course, but white noise doesn't elicit the strongest rate of this sort of error. People are constantly misidentifying things that are seen in a natural noisy env

      • by tmosley ( 996283 )
        The next one will recognize your gait.

        Crime is about o become completely impossible without the assistance of a specially trained AI assistant.
        • The next one will recognize your gait.

          Crime is about o become completely impossible without the assistance of a specially trained AI assistant.

          So put a small stone in one of your shoes and watch your gait change without you trying to "walk differently." Problem solved.

          • by tmosley ( 996283 )
            One more layer of neural net will see right through it.

            The problem is that in the not to distant future, it will start anticipating such ideas, and train itself to prevent confounding. Heaven is terrifying.
      • by Anonymous Coward

        Actually, in this well-documented [wikipedia.org] case of ceiling-walking bank robbery, wearing a rubber glove on one's head to look like a chicken was a very effective disguise.

      • by Tablizer ( 95088 )

        You try working spike-boots upside-down in a mask, bub

    • That's right out of Wallace & Gromit's "The Wrong Trousers".
  • by number17 ( 952777 ) on Tuesday February 17, 2015 @09:45PM (#49077609)
    The grocery store or ATM, with cameras all over the place, could do it by simply having a sign at the front of the store that says CCTV. They could record your picture at the register and associate it with a bank card or credit card. After 5 transactions they could guarantee that the person using that card has your face and is likely the owner. They could then flag how many times somebody else uses your card. They could track you throughout the store, like they do now but associated with an individual. Stores would have cameras at the entrances and exits. They would know how many people are currently in the store and who they are. They can't track though, I use cash, right?

    The grocery store subscribes to a third party a face recognition aggregator.

    At the beginning you'll have a shadow profile (#34950892). All it takes it one pump at a gas station, or taking cash out at the ATM to associate your face. About the only thing you could do is wear a disguise, but a different one each time.
    • The disguises and cash wouldn't be worth much in the way of anonymity if you were still carrying your cellphone.

    • Fortunately at least for the time being most public cameras are such bad quality images it isn't yet effective. How many times have you seen a robbery on the news and the guys face is not clear enough to get an ID? How many times does the parking lot have cameras, but you can't read the license plate? We have a limited window where we really aren't identified every minute of the day, but that will soon change. Even if a person doesn't participate in social media enough of our friends do that we are still un
      • Walmart already tracks your purchases and can figure out someone if a woman is pregnant base on buying patterns.

        Do you think they would say no to installing high res cameras in their stores to track what isles people walk down, what other products they stare at and for how long while deciding what to buy and associating that with a purchase?

        I can see them wanting to know what people look at and don't buy, so they can market specials on those products to them (mixed in with random specials, because people fr

        • The in store tracking can often be stopped by shutting off WiFi on your smart phone. Camera's can be involved there too, but they generally don't know who you are or link to previous visits without the WiFi bit. The purchases often are linked to using a rewards card or some such thing that gives them a way to link your purchases, not that every store doesn't have your credit card purchase history, but hopefully only for that store. I'd be curious if anyone has info on how much the credit card companies know

    • Secretly, for years, store employees have used advanced evolutionary facial recognition algorithms to identify customers. Where has the public outcry been?
      • If the store employees were somehow transferring their visual memory into a massive database then the outcry would be exactly the same.

    • ... All it takes is a facebook profile. They already do facial identification in the background.
      http://www.extremetech.com/ext... [extremetech.com]

  • http://arxiv.org/pdf/1412.1897v2.pdf

    • by godrik ( 1287354 )

      I don't see the practical relevance of this? You can not walk through an airport with a scrambled face. So the images the camera will get are "regular" imaegs. Sure you can generate ridiculous images that triggers false positive. But these images will probably not be fed to an actual system.

    • by serviscope_minor ( 664417 ) on Tuesday February 17, 2015 @10:21PM (#49077777) Journal

      Debunked?

      They're a machine learning algorithm. All such algorithms do is place a fancy decision boundary in a high dimensional space. DnNs do a decent job for certain classes of problem. Far away from the training data, the boundary is not useful, but that's the same with all algorithms pretty much.

      So no. They haven't been debunked.

  • by Anonymous Coward on Tuesday February 17, 2015 @10:17PM (#49077765)

    Very much anecdotal, but here goes anyway - a little while back, I found a recipe for cow tongue that seemed intriguing. If I had eaten it before I couldn't recall, at least I hadn't prepared it myself. So off to the butcher's I was, as this is not found in every shop. The tongues they had on display there seemed very tiny (in retrospect, they must have been veal tongues), so I said "give me the largest tongue you have". As the saying goes, "you should be careful what you wish for" - what I ended up with was a monster, something like over 1.3kg (nearly three pounds). I really didn't need that much, but all I could do was to say thanks and go home with my prey.

    As I laid it on my cutting board, pretty much filling it entirely, it looked at the same time so awesome and gruesome that I had to take a photo of it (not a food blogger, or a blogger of any kind, I just had to document it). And to share the experience, I sent it to a friend via Hangouts. Now, as she uses Hangouts from the GMail web interface, the images are not visible inline but are Google+ links. So she clicks the link.

    ...and G+ helpfully asks her "Is this xxxxx?" (xxxxx == her name) While people are, rightfully, concerned whether companies such as Google know too much about their lives, at least when it comes to Google and facial recognition, they have a long way to go.

  • When there is a competition to test solutions, do they call it a "face off" or a "face face off"?

  • The facial recognition software thinks all Asians are the same guy.
  • So... (Score:5, Funny)

    by jtownatpunk.net ( 245670 ) on Tuesday February 17, 2015 @11:32PM (#49078125)
    Can I finally automatically tag the performers in my porn collection? I'm asking for a friend.
    • by Anonymous Coward

      nope, this only does face recognition

  • by Torp ( 199297 ) on Wednesday February 18, 2015 @12:49AM (#49078317)

    I didn't read the article, of course, but the summary sounds like they're doing face *detection* not recognition.
    Detection: find which portions of an image are faces.
    Recognition: compare to a database of faces and find out whose face it is.
    First is way easier than the other.

  • by Anonymous Coward

    I wish that interesting developments in algorithms such as this could be discussed without resorting to cynicism (as in, how they'll be used by the NSA to breach our privacy and etc). Yes they're valid concerns but my God, what point is there in enjoying advancements in technology if you're going to see the downsides in everything? I long for a simpler time when we didn't need to worry about such BS.

    • I long for a simpler time when we didn't need to worry about such BS.

      One of the main reasons we have so many problems with privacy and security these days is precisely because of those simpler times when nobody really worried about the implications of various technologies. We're much better off being cynical.

  • by MrL0G1C ( 867445 ) on Wednesday February 18, 2015 @09:26AM (#49079085) Journal

    Facial recognition mostly gets used for all of the wrong reasons, Facebook tracking, illegal police tracking etc.

    'photos of innocent people have been retained in contempt of an explicit order from the court to remove them' - 18million by police [techdirt.com]

    Facebook's new face recognition policy astonishes German privacy regulator [pcworld.com]

    And what about people who don't have Facebook accounts, does Facebook allow 'tagging' of their faces?, I'm already annoyed by Facebooks obvious data collection on me as shown by the fact I get email from them telling me who my friends are and inviting me to join.

  • This plus drones with missiles = aimbots in RL
  • Have I missed something?

    I've always believed algorithms and neural networks to be essentially opposites to each other.

    Algorithms are blocks of code that handles a predefined task. Classic example: quicksort vs bubblesort

    Neural networks are a black box of systems that are trained with input until they produce the output you want. Further, even when it is working, you won't truly know what is happening internally, and you're only hope of knowing that it works is throwing a ridiculous amount of inputs at it

    • I've always believed algorithms and neural networks to be essentially opposites to each other.

      I think you're mixing up two different levels of abstraction.

      Algorithms are blocks of code that handles a predefined task.

      Indeed, and the NN algorithms describe a predefined task that can be summarized as "train and operate a neural network". That's one level of abstraction. Once trained and operating according to the algorithm, then the NN proceeds to do the tasks it is meant for. There is no algorithm that the NN follows at this level of abstraction -- there is only the algorithm for how the NN itself operates, not for the specific task that NN is being used for.

  • What's more, their algorithm is significantly better at spotting faces when upside down, something other approaches haven't perfected.

    Very usefull if you want your system to work in Australia!

  • OTOH, what these networks have actually learned can be eye opening [youtube.com].

  • Don't those guys on the NCIS TV show do this all the time?

"Truth never comes into the world but like a bastard, to the ignominy of him that brought her birth." -- Milton

Working...