By Xeni Jardin at 11:03 am Wed, May 9, 2012
Kyle McDonald and Aram Bartholl of fffff.at teach you how to avoid being "auto-recognized" in photographs: surveillance shots, party pictures on Facebook, you name it. The short version: tilt yer head.
This will work until the algorithms are updated to compensate for it.
I was going to say the same thing… don’t rely on this. You will be recognized shortly. Even if not now, your pictures will still be up when the algorithms have been updated. Only way to guarantee avoiding facial recognition in a pic is a mask. And even then I wouldn’t be surprised if body shape/posture recognition is developed in the future.
I wonder if those cameras pick up IR. then you can wear a simple headband that will blind cameras and be invisible to people.
Pittsburgh Pattern Recognition (recently acquired by Google) had algorithms that worked pretty well against tilt.
I once posted a picture to FB of myself in shorts. FB wanted to know who my knee was. So…there’s that.
iPhoto didn’t spot me right in front of the camera, but had some NYC street weirdo off in the distance pegged as me instead. Funny aberrations.
Man, Berlin looks more and more like New York, people wise.
(Not that that’s a bad thing,)
Man, you weren’t kidding about “the short version.” I was expecting something in those two and a half minutes to explain why this should work. I can’t believe it would, nor would it be remotely difficult to defeat if someone in the surveillance office started wondering who all those “unidentifiable” people walking around are. Identifying eyes as eyes isn’t difficult, particularly for any camera used for facial recognition. Rotating the image to put those eyes in line with each other is trivial. If the software has a bug that doesn’t allow for crooked heads, and we know about it, can we really expect it won’t be patched by, like, last Wednesday?
In any case, even if it works on Facebook, how many surveillance cameras are mounted at eye level? Outside the movies, have they never been able to identify people using those off-axis cameras?
They also make it look like their voice is coming out before their mouth moves. Amazing trick!
Fucking tilted head hipsters.
Wait! what is going ON in that video? I see people with faces… and then their heads just disappear, and the space above their necks is replaced with an amorphous unrecognisable ‘blob’. BUT I CAN STILL HEAR THEM TALKING! Then they say they’re tilting their heads back, and suddenly I can see them again. Give it to me straight-up Dr. Sacks … does that mean … does that mean I’m a robot??
Nexus 6, I suspect.
Or you can simply turn off facial recognition in your facebook privacy settings.
Facebook doesn’t protect you well against your idiot friends posting pictures of you. Furthermore, Facebook doesn’t control all uses of the picture, even if it was posted there originally.
Which will only make YOU not know who is identified. Every other picture with is still identified and of course, clicking out the “show it on this pic” option doesn’t mean that Facebook doesn’t identify you. They yust don’t show it to you.
Or you can not use Facebook :)
That helps you against your bastard friend with a camera and a tag-ready mouse exactly how?
An out-of-sync, poorly shot, overly-ambitously edited, 2 and a half minute video with about 5 seconds of (questionably) useful information. Gee thanks.
New to European cinema, eh?
At least as far as the 3DS is concerned, when it is trying to detect faces I apparently have no face. I’m assuming that the combination of my facial hair and glasses is enough to throw its face detection algorithm off. I don’t have enough data to state whether that applies for other algorithms, but at least I can take some comfort in knowing that some systems won’t recognise my face simply because they don’t see me as having one.
Sound to me like they’re just trying to make people walk around looking stupid.
Unlike these guys: http://cvdazzle.com/
Much cooler solution. Also because we finally get to look like we’re living in the future.
Old joke from the coldwar (un-known if true or not)
Us spends 20 billion$ to build stealth fighter that has radar cross section of seagull
Russia spends 200$ upgrading software to spot seagulls moving at 200mph
Adaption, it’s not just a movie.
Spoof the software this way, and you’ll soon find it can counter that.
You’ll be doing dumb stuff thinking your safe, and are far from it.
That out of sync business ruined any message they may have had for me…
so the reason this works is because the computers don’t know how to find faces where the rotation is beyond a certain arbitrarily defined point (it’s more computationally expensive and of relatively little value for the algorithm to be able to locate faces under all possible rotations).
that said, it’s certainly possible to widen the rotational range that the algorithms can deal with so i wouldn’t rely on it.
frankly, avoiding facial recognition is not hard or complex and it doesn’t even have to draw attention. they’re right that masks obscure the face, but you don’t have to obscure the entire face, only an important part. the eyes happen to be an important part, and sunglasses can obscure those just fine.
I dunno. There are some people whose eyes aren’t their most distinguishing characteristic.
perhaps not their most distinguishing characteristic in terms of telling one person from another, but in terms of telling a face from a door or potted plant, eyes are very important.
I ask people to not tag me in photos. Also, if I do any tagging, I do it incorrectly. Degrade the signal to noise…
I don’t like the general idea of showing up anywhere, picturewise, and not being able to control it. Apart from that, the video was shot ten minutes away from my home.. ^^ my home area looks weird through a lens.
I work in face recognition, and this is pretty ridiculous. They are talking about “roll” rotation, which is the easiest to compensate for because it is a simple image rotation. All the information is there, it is just rotated. The default for most algorithms is +/- 15 degrees, but that isn’t a technical limitation, it is just a convenience for processing speed. Any algorithm can be configured to check +/- 180 degrees quite easily. Algorithms have a much harder time on “pitch” and “yaw” because now you are hiding facial missing information. A better solution would be to just look down when you are near a camera. Because most cameras are mounted high, they won’t be able to see much of your face at all.
CVDazzle has gotten a lot of press, but it hasn’t been significant to me in my experiments. See: http://www.forbes.com/sites/andygreenberg/2010/09/27/how-to-hide-your-face-from-big-brother-try-sunglasses/
A mask, of course, will work quite well. Another simple solution is dark sunglasses.
A better solution would be to just look down when you are near a camera. Because most cameras are mounted high, they won’t be able to see much of your face at all.
Revolution of the Shoegazers!
They had this figured out a _long_ time ago…
Tilt of the mentioned 15 degrees is not even a problem for my Panasonic DMC-ZS20 point and shoot camera, which recognizes faces of people stored ahead of time and puts their names on the screen. I tried different tilt angles, and recognition failed closer to 30 degrees of tilt. Interestingly, adding images tilted beyond 30 degrees to the comparison registry did not help, indicating that the basic idea of out-of-level beyond a certain roll rotation causing issues with recognition is sound, even if the mentioned angle was too small. But if my little camera has enough processing power to handle close to 30 degrees, as @AlexKilpatrick mentioned, I would imagine law enforcement has enough processing power to handle arbitrarily high roll angles, in case perps start strolling around with tilted noggins in great numbers.
Pfft. The real secret’s in Chris Morris’ film Four Lions where the would-be jihadi shakes his head continually so it comes out “blurry” on the CCTV picture.
I doubt this works unless the algorithms are horrible or haven’t been trained much.
I use Google’s Picasa and the recognition is really quite good. It sometimes correctly identifies people even if their entire face is not in the shot. It’s nowhere near 100% but it helps a lot. Without it I’d never have gotten the hundreds of thousands of faces in my tens of thousands of personal photos indexed.
Submit a tip
The rules you agree to by using this website.
Who will be eaten first?
Jason Weisberger, Publisher
Ken Snider, Sysadmin