Re:A Global Database for Facial Recognition
by evandavid on Wednesday January 26, 2011 @08:32PM
(#2146153)
I worked in Biometrics for a few years. I was on a fingerprint team, but I was part a wider team that was also responsible for forensic-quality facial recognition technology (a world leader). To put it bluntly, the technology is still very, very, very weak. When it works, it's really impressive, but that's normally during a demo from a nicely curated database. There are some facial technologies that can "work" from a distance (and they are used in airports already), but the success rate is low and the original image still needs to be of decent fidelity. Often times the results from the automated matching will be shortlisted for comparison by human operators. Maybe I was blinded by the awesomeness of fingerprint technology (still blows my mind thinking about it).
Combine this general lack of automated matching awesomeness with the fact that people age, can wear glasses or a beard, etc, and we are still many years away from this kind of capability.
Still, it's always fun to think about the possibilities that biometrics can and will provide.
Re:A Global Database for Facial Recognition
by liuliu on Wednesday January 26, 2011 @09:35PM
(#2146337)
Sometimes I have my doubt on how current facial recognition technology gonna perform on tens of millions samples. Admittedly, we have become quite good at the scale of tens of thousands real-world data (see Labeled Face in Wild project), but how reliable it really is on large scale? My friend circle (everyone I have met in life for a small period of time) probably only consists of few thousands people, but I can remember many people who are distantly related have very similar faces. My main concern is, the facial variations between people may not be as large as we have imagined once we are on tens of millions scale.
Here is an idea to validate my concern:
1. Crawl over all photos on Facebook and run a frontal-face detector with it;
2. Use current state-of-art methods (attribute based method or hybrid methods) to get several pairs of similar faces (highest score ones) of different people;
3. Use the same method as in step 2 to get several pairs of similar faces of same person with roughly the same score as in the step 2;
4. Have human volunteers to blindly judge which pairs of person are in fact the same one given only a small region of picture (e.g. face should occupy 20% of the total area);
5. Check the result in step 4 with ground truth data;
Re:A Global Database for Facial Recognition
by apu on Wednesday January 26, 2011 @10:24PM
(#2146429)
The face recognition problem is very tough and we are nowhere close to solving it for most real problems. The results on Labeled Faces in the Wild (LFW) [1] show that even on the easier "verification" problem -- "are these two images of the same person?" -- the best performance is under 90%.
To figure out how this translates to recognition -- "who is this person?" -- you can roughly take the verification rate V and number of different people in your database N and get a recognition rate of roughly V^sqrt(N). (This is a very very rough estimate!)
So with 1 person in your database, your rate is just V = ~90%. With 100 people, it's already down to 35%. And so on...
(I'm the author of one of the best methods on LFW right now -- the attribute-based one [2].)
Re:A Global Database for Facial Recognition
by liuliu on Wednesday January 26, 2011 @10:51PM
(#2146497)
I am eager to replicate your attribute method but haven't got much time to do so. Very excited to know that you are on Hacker News too. My impression is that "verification" problem is already a leap forward comparing to "identification" problem which consists of fewer people and many samples. The problem space is much interesting now than before with the astronomically number of digital photos available on the Internet (considering that for about 3 years, the "verification" result are improved from 69% to around 85%). But my main concern is that our human being may not be good at "large-scale" face recognition problem after all. If that were true, the fundamental of computer-based face recognition problem would be flawed.
Re:A Global Database for Facial Recognition
by apu on Wednesday January 26, 2011 @11:50PM
(#2146631)
We have results in our face verification paper showing exactly how well people do on the LFW verification task. The answer is: very good, at 99.2%. However, we also show that much of this is coming from the background and context in the images in LFW. The fact that those images were originally gathered from Yahoo News means that many images are taken at the same event -- with the same background and same clothes, etc., and people can use this to make good guesses, even without seeing the person's face!
I think the vision community is thus looking for how to exploit context effectively. There have been a number of interesting approaches recently, from looking at "things" vs "stuff", to using additional information such as transcripts and captions, etc.
And of course, there are still lots of people looking at how to better solve the basic recognition problem itself.
Re:A Global Database for Facial Recognition
by uptown on Wednesday January 26, 2011 @09:03PM
(#2146250)
Facial-recognition alone is probably a ways off, but if you were able to combine a number of digital hints you'd be able to give your system a tremendous boost. Think about everything we do that's logged into a database.
cell phone location, credit-card purchases, mass-transit card swipes (and cameras by those turnstiles), future location-specific purchases (airline tickets, concert tickets, etc.), EZ-Pass for your car, IP address of your web browser, location-tagged tweets, foursquare check-in, historical matches based on a routine, and on and on and on.
You may not be able to easily match a random face against your entire database ... but if you were able to combine all of those elements together you could almost say where someone was more times than not without even having an image of the person. Right now only the government could get access to everything listed above ... but if the trend of over-sharing and openness continues it's not a stretch to imagine this being heavily commercialized. Depending on how many of their services you use, Google probably has a pretty good idea where a large percentage of their users are at any given time. Combine that with the natural evolution of StreetView (live video) and you've got your product.
Re:A Global Database for Facial Recognition
by trotsky on Wednesday January 26, 2011 @08:23PM
(#2146135)
As I understand it facial recognition only works that way on tv. Facial recognition is still hugely computationally expensive. In the real world it's used by security services by having a good picture of known bad guy x and then having the system watch cctv in a couple of zones to try to spot him. Even that takes hella cpu time, but it's still matching only a small fraction of the faces... thousands vs. Hundreds of millions.
Re:A Global Database for Facial Recognition
by raphar on Wednesday January 26, 2011 @09:31PM
(#2146326)
I'm sure that if the gambling industry think that this idea is useful and feasible, there is a system using such database already installed on their casinos.
Re:A Global Database for Facial Recognition
by Tichy on Thursday January 27, 2011 @03:33AM
(#2147127)
I keep reading "crawl Facebook". How viable is that kind of thing? I could think of various ways to use the data (like I proposed a friend I could create a service that would help him identify the woman he did not approach - as a bonus, that project might boost privacy awareness). But I assumed that crawling Facebook would not be appreciated by Facebook, and also easily prevented by Facebook. Since there are 500 million profiles to crawl, a lot of different IP addresses for the crawler would be required.
Also, how legal is it to actually use data from such scrapes?
Re:A Global Database for Facial Recognition
by greendestiny on Wednesday January 26, 2011 @07:56PM
(#2146053)
Yeah I've thought about it before, and I suspect Facebook would likely object on copyright grounds. You could argue that it's simply an index to a visual search engine.
I wouldn't be surprised if people had done this though, just not made it publicly available.
Re:A Global Database for Facial Recognition
by spoiledtechie on Thursday January 27, 2011 @01:39AM
(#2146874)
Its already been done. By private companies selling to the Highest bidder, mainly the Govt. How do I know, cause I have been watching my coworkers working on the project for the past year. Its just not public yet...
Re:A Global Database for Facial Recognition
by 51Cards on Wednesday January 26, 2011 @08:28PM
(#2146146)
There may not be an expectation of privacy in public, but I suspect it could be argued there is an expectation of anonymity. The tech is a long way from this being a reality, but the thought of it is beyond scary.
Re:A Global Database for Facial Recognition
by ricaurte on Wednesday January 26, 2011 @11:47PM
(#2146625)
Face.com seems to be essentially going down that path. Now whether they store all of the faceprints or not, I don't know. They also have a developer API that is in alpha right now.
Combine this general lack of automated matching awesomeness with the fact that people age, can wear glasses or a beard, etc, and we are still many years away from this kind of capability.
Still, it's always fun to think about the possibilities that biometrics can and will provide.
Quite creepy.
Here is an idea to validate my concern:
1. Crawl over all photos on Facebook and run a frontal-face detector with it;
2. Use current state-of-art methods (attribute based method or hybrid methods) to get several pairs of similar faces (highest score ones) of different people;
3. Use the same method as in step 2 to get several pairs of similar faces of same person with roughly the same score as in the step 2;
4. Have human volunteers to blindly judge which pairs of person are in fact the same one given only a small region of picture (e.g. face should occupy 20% of the total area);
5. Check the result in step 4 with ground truth data;
To figure out how this translates to recognition -- "who is this person?" -- you can roughly take the verification rate V and number of different people in your database N and get a recognition rate of roughly V^sqrt(N). (This is a very very rough estimate!)
So with 1 person in your database, your rate is just V = ~90%. With 100 people, it's already down to 35%. And so on...
(I'm the author of one of the best methods on LFW right now -- the attribute-based one [2].)
[1] http://vis-www.cs.umass.edu/lfw/
[2] http://www.cs.columbia.edu/CAVE/projects/faceverification/
I think the vision community is thus looking for how to exploit context effectively. There have been a number of interesting approaches recently, from looking at "things" vs "stuff", to using additional information such as transcripts and captions, etc.
And of course, there are still lots of people looking at how to better solve the basic recognition problem itself.
cell phone location, credit-card purchases, mass-transit card swipes (and cameras by those turnstiles), future location-specific purchases (airline tickets, concert tickets, etc.), EZ-Pass for your car, IP address of your web browser, location-tagged tweets, foursquare check-in, historical matches based on a routine, and on and on and on.
You may not be able to easily match a random face against your entire database ... but if you were able to combine all of those elements together you could almost say where someone was more times than not without even having an image of the person. Right now only the government could get access to everything listed above ... but if the trend of over-sharing and openness continues it's not a stretch to imagine this being heavily commercialized. Depending on how many of their services you use, Google probably has a pretty good idea where a large percentage of their users are at any given time. Combine that with the natural evolution of StreetView (live video) and you've got your product.
Also, how legal is it to actually use data from such scrapes?
I wouldn't be surprised if people had done this though, just not made it publicly available.