Ever have that feeling that you’ve seen a face before, but just can’t place it? Is it someone you knew in high school, or someone you saw featured on TV as a crazed serial killer? This won’t be a problem in the future, thanks to the folks at Australian think tank NICTA – with assistance from GPUs. In the “Enabling Large-Scale …
Does anyone believe this?
"Using a Tesla blade containing four GPUs, NICTA can match a single face back to a database of two million images in less than a second"
Yes I do believe this...
I'm assuming they are not storing images which they compare, but instead storing and comparing Biometrics they previously extracted from photos. In which case that's vastly less data to compare.
Furthermore I'm assuming they are describing a so called "S2050/S2070 1U GPU Computing System" i.e.
This is capable of 4TFlops (from a total of 1792 cores). Plus its got 593.6GB of memory bandwidth. So 2 million results per second sounds very possible.
So yes, I do believe its possible, I just didn't realise how close we were to such massive scale facial recognition technology until now. (I hadn't put the two ideas together, so yes it does sound very feasible).
Re: Yes I do believe this...
They may not even have to base it on pre-calculated bio-metrics... Given the memory bandwidth, it may simply be comparing the photos on the fly.
Assuming each photo was around the 5KB mark, if they loaded 2 million photos into RAM, it would still only need 2GB to hold it all. Hardly a massive storage target these days.
Amazing and somewhat shocking that what was once unthinkable, is now starting to become possible...
@"Using a Tesla blade containing four GPUs, NICTA can match a single face back to a database of two million images in less than a second – which is pretty damn speedy."
Ok very quick very rough estimate based on current technology: … The entire population of the UK would be a database of 60 million images which will take a server of just 30 Tesla blades (each with four GPUs), to once again find one face per second. That's not fast enough for real time video surveillance, but its fast enough for a central police super computer to lookup mug shots on requests from the police. (Amazingly even assuming 10 photos per person, this is still becoming a practical system).
Another way to think of it is based on this technology only 3000 Tesla blades could therefore handle roughly 100 face search requests per second where an answer was returned in just 1 second on average. So even if they had to wait 10 seconds during peak busy periods (Friday nights?) that's still a very likely system for them to try to build.
Now how usable that is and how many false positives and negatives degrade it I don't know, (although its usability will be improved by comparing the returned search data with any ID on the person) plus even just a name and address will allow a much more detailed search.
Given this kind of advance in facial recognition speeds, I can see a time not long from now when every time police speak to anyone, they will photo them to amend comments to their report included with the photo. Then other police will try to recall these comments from their colleagues in the same way domestic extremist labels are used now, but the labels about people's behaviour when questioned by police will be able to become far more subtle. (e.g. This one spoke back to me, this one doesn't listen to everything I tell them). It will mean we are moving quickly towards a world where we all have police records, simply due to accumulating data, so not just police records for criminals, as there is ever more reasons (the people in power will find) to photo us all. (This is already what is happening with the Domestic Extremist label that police already use, plus the way the police photo protest meetings etc..).
So I really do find it amazing and somewhat shocking that what was once unthinkable, is now starting to become possible.
Before anyone tries to patent this idea...
Check out Japanese Anime "Eden of the East".
If Moore's law holds true... in about 10 years your phone will be able to do this right and still let you surf the net for porn at the same time. :-)
Ok,I object to this on Civil Liberties grounds...
... but it would be incredibly useful for people like me who find it extremely hard to remember names and faces, next time I see someone, a quick zap with the camera and it can pop up their name and other details to save me having to say "Sorry, I can't remember who you are"!
Smile (for the camera!)
If you've done nothing wrong, what have you got to hide? Goes the question asked over and over again by the "I'm happy and like the idea of being scrutinised every minute of the day, it makes me feel safe" brigade. Well, the simple answer is that no matter how innocent. blameless and generally bound for sainthood you might be - or think you are, there's some scrote out there who looks just like you.
If the mugshot of that person shows every blocked pore, every chipped tooth and nasal hair, then a reasonably observant person can, when presented with your phizzog in person and a high resolution print of said photo, tell that you're not the same people. However if the pixel-limited, poorly lit, motion-blurred and partially shadowed glimpse of someone doing something bad correlates on someones computer with the glum, washed out passport photo [ BTW: smile on your PP photo and it'll be rejected - they like their travelers to look oppressed and down-trodden] you submitted some years ago then it's chokey for you, sonny Jim. At least until a real person gets to review the "evidence" after some days of inconvenience and scurrilous rumours amongst the neighbours, when your front door was smashed down at 5:00 a.m. by the anti-<whatever> task force.
It is pretty obvious that you can't make data out of nothing (except in some cheezy cop shows, when you can zoom in on a mobile phone pic. practically down to the quantum level). Most people's eyes are 60 + a bit mm from pupil to pupil. If you have 80 or 90px between eyeballs, you get sub-millimetre resolution. Drop that down to 12px and your resolution (even using sub-pixel interpolation) is going to match maybe one in every ten people just based on their eyes. Add in all the other facial recognition points and you'll find everyone in the UK matches maybe a few hundred or a couple of thousand other, non-burka wearing, individuals - even under the best conditions. You're effectively taking part in an involuntary identity parade every time you're subjected to this sort of facial recognition technology. I just hope you can account for your movements at all times - you might just need to, one day.
Imagine we tried it here...
Let's say 10% of the people on a 1.5 million person watchlist make one journey through British airports in a year and that 95% of these are recognised by the system. Let's say 230 million other passengers pass through the system with a 1% false positive rate. That is 142,500 genuine recognitions out of 2,300,000 + 142,500 recognitions in total, ie: 5.8%. Even if 90% of the false positives can be quickly cleared by checking their documents, that still leaves 230,000 false positives plus 142,500 people from the watchlist to be investigated more closely. That is 372,500 people to investigate, only 38% of whom were on the watchlist. Spread evenly over the year, that is 1,020 new cases per day, only 390 of them from the watchlist. That ought to take the pressure off homegrown terrorists!
No figure given on false positives?
I simply do not believe that this would be possible without a massive number of false positives (and negatives). If they come to rely on it there will be huge problems.
Oh, and "passport photos. The subject is usually standing right in front of the camera, smiling nicely" - when did you last get a passport renewed? Smiling and looking happy is specifically forbidden, citizen.
Yep, you're right
It's been a while since I've renewed my passport photo, but I do remember that 'no smiling' rule. However, is this the case for every country? Couldn't there be a govt. out there that realizes that their country is such a stinking hole that any rational citizen should be grinning like an idiot at the possibility of leaving it, if only for a short trip?
Let's see what NICTA actually say
From: David Moss
Sent: 28 September 2010 10:14
To: Professor Brian C. Lovell
Cc: Dan Olds; John Lettice; Joe Fay
Subject: NICTA say real-time matching of CCTV to mugshots is here today -- or do they?
Dear Professor Lovell
An article published by Dan Olds of Gabriel Consulting Group <http://gabrielconsultinggroup.com/> , Oregon, was published in The Register <http://www.theregister.co.uk/2010/09/27/facial_recognition_nicta/> magazine asserting that NICTA claim that low resolution CCTV images and/or camera stills (it's not clear which) can be reliably and quickly matched to a database/gallery of mugshots. The article makes no distinction between one-to-one and one-to-many matching and provides no figures for false positives and false negatives.
The article says, for example, "The tool [unnamed] is also very flexible and lightweight enough to run on a smart phone. So if, for example, a liquor store was robbed, the officers on the scene could snap a picture of the thief off of the surveillance system monitor (even with their typically crappy quality) and search a mug shot database there and then".
A quick trawl reveals this NICTA article <http://www.nicta.com.au/__data/assets/pdf_file/0009/14958/Smart_Cameras_Enabling_Automated_Face_Recognition_in_the_Crowd_for_Intelligent_Surveillance_System.pdf> , which states that "Ideally, a robust face recognition system would be able to recognize faces regardless of the face’s expression, angle, features and lighting conditions". That falls a long way short of the claim made by Mr Olds.
There is a reference to another NICTA article <http://www.nicta.com.au/__data/assets/pdf_file/0009/14949/Reliable_Face_Recognition_for_Intelligent_CCTV.pdf> which says that "With passport quality photographs, current face recognition technologies can approach 95% recognition accuracy. Yet trials show that performance drops to only 10 to 20% when there are significant changes in lighting, pose, and facial expression". No explanation there of "95% recognition accuracy" and no references to support it but, again, we seem to be a long way from Mr Olds's position.
I believe that the readers of The Register may be being misled by Mr Olds's article and that he may possibly be misreporting NICTA's position.
Historically, facial recognition has always failed in large-scale field trials and the computer simulations conducted by the likes of NIST do nothing to increase confidence in facial recognition technology. For what it's worth, I am all in favour of research into biometrics but I am against governments wasting taxpayers' money <http://dematerialisedid.com/Register/regBiometrics.pdf> deploying technology that costs more than the limited benefits it can deliver.
The NICTA references led me to you. May I ask you, is it NICTA's position that one-to-many matching from CCTV images and camera stills is now, today, possible in near-real time?
You don't have to answer, obviously, and any answer would be likely to be published.
But it really would be interesting to know if there are any large-scale field trial results anywhere in the world which actually support any of the wilder claims <http://dematerialisedid.com/PressRelease22.html> made for the reliability of today's facial recognition technology.
Settle down, D Moss Esq.
First of all, this was a blog entry, not a research report. It was intended to highlight an interesting interview and session at the GPU Conference. NICTA's facial recognition technology seems to be quite a bit ahead of other technology that I've seen out there.
I'm not an expert in this field and do not present myself as such, but I do have a smattering of technical background. I believe that what they presented at the show fairly represented what they can do now and is a preview of what the technology will be able to do in the future.
It seems that you have a bit of a problem with their approach and results, which is fine. But I'm not overly pleased with your saying that my article "misled" or was filled with "wild claims.” This isn't the kind of behavior that I'd associate with an Esquire, and it saddens me. Here in America, we expect more from guys who append ‘Esq.’ to their names.
It's probably relevant for our readers to know that you have your own agenda here. You have some sort of proposal dealing with some identity system based on mobile phones or some crap… I found it online, but lost interest pretty quickly.
I'd suggest that you take up your concerns with NICTA directly rather than use open letters posted in online publications. They seemed like a pretty good bunch of guys, and they'll probably respond if you approach them in the right way. You know, like someone who actually wants to learn something – not like a blowhard gadfly.
Dan Olds: "I'd suggest that you take up your concerns with NICTA directly"
My email is addressed to Professor Brian C Lovell, with cc's to Dan Olds and others.
Professor Lovell was not chosen at random, he is a Research Leader at NICTA , he is one of the authors of one of the NICTA papers I quote from and he holds 24 distinguished academic positions listed on his NICTA web page . He was chosen because he looks as though he should know what he's talking about and could talk with some authority on behalf of NICTA.
- Review We have a winner! Fresh Linux Mint 17.1 – hands down the best
- Vid Antarctic ice THICKER than first feared – penguin-bot boffins
- Antique Code Show World of Warcraft then and now: From Orcs and Humans to Warlords of Draenor
- iPhone sales set to PLUMMET: Bleak times ahead for Apple
- HTML5 vs native: Harry Coder and the mudblood mobile app princes