Slashdot Mirror
Face Recognition — Clever Or Just Plain Creepy?
Simson writes "Beth Rosenberg and I published a fun story today about our experiences with the new face recognition that's built into both iPhoto '09 and Google's new Picasa system. The skinny: iPhoto is fun, Google is creepy. The real difference, we think, is that iPhoto runs on your system and has you name people with your 'friendly' names. Picasa, on the other hand, runs on Google's servers and has you identify everybody with their email addresses. Of course, email addresses are unique and can be cross-correlated between different users. And then, even more disturbing, after you've tagged all your friends and family, Google tries to get you to tag all of the strangers in your photos. Ick."
They are trying to covertly establish a database for their next big project; Google People Search. With just a name you can find a person's address, email address, phone number and what they have searched for in the last three years.
What exactly did everyone think "Don't be Evil" would mean once the company went public, grew up and grew larger?
Not that this is necessarily anything premeditated and sinister, but notice how thinking through whether something might seem weird or discomfiting isn't at the top of the list?
Taking away the facial recognition technology, it's not that much difference than facebook. A friend takes a photo of me somewhere, sticks it on their facebook profile, labels me in the picture, and links it to my facebook profile. Then your pictures can be searched.
Given enough labeled pictures of me, one could run it through a facial recognition system. It would have the same applications, without the initial creepy factor.
Talking about facebook, I guess soon people will not need to label you. Facebook will label you automatically. Recognition error rates can be reduced by making sure you are in the same circle of friends.
Assuming that the uploaded pictures also contain the proper EXIF data, then Google will also know exactly when was the picture taken. If you they can also figure out the location the picture was taken on (perhaps as a tagging feature connected to Google Maps?), then they'll be able to track people - where and when they were, and in whose company. They could even extend the concept to try to combine pictures of the same event from different albums into a massive "super-album" of the event, even if the owners of the photographs never found out about each other.
It's all creepy to me. The fact is, it's pervasive and very difficult to opt out of as current social norms exist. Even if you don't have a gmail account, if you have even a small circle of friends there is a high chance of someone else having a gmail account that you have sent mail to, which puts your email in that circle of friends. If someone else in the same circle of friends has your picture and labels it, that would be enough to reliably link your email, first name, last name and face together. (Your friends would be identifiable as a cluster.)
With the above and a sample of your writing, there is a good chance that you have a statistically improbable phrase or two (or vocabulary) that is going to identify you elsewhere on the net.
Have enough cameras in a given country, and you can build up a database of people and locations they have been to, updating in real time. Those whose faces haven't yet been identified will relatively easily be able to be associated with the groups of people they associate with (enough times at nearby camera locations with a given person at the same time, with extra weight if those other people are there at the same times, coupled with cell phone location information), and their domicile located to within the nearest camera. In fact, just correlate the cell phone location info with the face from the camera - if they have a phone you have a match. Remember there is also the database of passports (with photos) that can be assumed scanned, and nicely labeled high-school graduation photos for all potentially subversive people coming of age. And those unidentified faces might be driving a car which will have a license plate, again, traceable to a database of names and addresses.
At this point the number of unidentified people should be small enough so that there are only a relative handful of people who are just unidentified faces. These people will be probably be high value enough on average to make it worthwhile to find out who they are. It would be relatively inexpensive to obtain their identity - identify from the database the scheduled activities they keep, send an undercover vehicle there to stake them out at probable times, when the camera gives a positive, trail them home to an address etc. You'll get at least an alias if not a name, an address, a face, and a likely circle of friends.
If they can recognize faces they should be able to recognize ethnicity (if you can recognize a face and an ethnicity from a photograph, so can a machine). The facial measurements and last names will form a cluster. A scan through the last names will identify the ethnicity probably within the first few entries or so.
If I can think it, google/NSA has smarter people than me working for them and they will have done that and more.
The only way to opt out is to live as a hermit or with similarly google avoiding hermits. Maybe not even possible. It seems harmless enough now, but the moment the rulers are actually fearful (e.g. if there was a large enough depression, people out of work started rioting in sufficient numbers or with arms), you can bet that there will be unmarked vans going around the city in the night picking up people with their "SubversiveRank (TM)" above an arbitrary threshold with a one-way ticket to either a slave labor camp or an unmarked grave.
Who exactly does google's "Don't be Evil" motto apply to? It makes MUCH more sense if it is externally directed.
If I have seen further it is by stealing the Intellectual Property of giants.
... Google should get. It wants you to name all those people? That's Sergey Brin and Larry Page. All of them. Google wants email addresses? Get a gmail account, tag them all with it, spoof yourself with it, and then surf a couple dozen porn sites and post to a bunch of usenet groups. Google wants mail? Give it to them.
"I may be synthetic, but I'm not stupid." -- Bishop 341-B
NSA and the Department of Homeland Insecurity likely already have face recognition software trawling websites including social network websites recognizing the same people popping up in photographs all over the world. I doubt it's effective in practice, if they have this, but in theory, this would be the technique to be able to 'search' for say, a suspected terrorist, drawing down shots of faces from all over the world. Someones holiday snap of a crowd in some city. pulled down from flickr. may put a pin in the map as far as tracking a known suspect goes. Nevermind what realtime access to urban CCTV that seems to be popping up in many cities all over the world.
^^^You see what I did there right... I fixed it for myself, i put 'IN' in front of 'security'.
After logging in slashdot still does not take you back to the page you were on. It's been that way for 20 years.
Thoughts: A "fuzzing" of an image, saved down to a couple of SHA hashes and an MD5 would give you a "close match" system which you could then recognise a lot easier.
So, light colour variations in cheeks (for example) are removed (blended out) of the image, hashes are taken, close matches are processed harder for tighter possibilities.
Perhaps hashing isn't the right answer, maybe we could look at pixel-area-colours and match from there? Too many thoughts, too late in the evening.
Warning: If you are a creepy government organisation, or Google, do not read this post.
Me failed English...
FreeBSD over Linux. If my comments seem odd, this may explain...
Fractal compression would do the job if it were not encumbered with a pile of patents issued a couple of decades ago.
And did you exchange a walk on part in the war for a lead role in a cage? - Pink Floyd.
Actually that would be very close to what I was thinking. Remember that guy from a few weeks (months?) back who created the Mona Lisa ( http://developers.slashdot.org/article.pl?sid=08/12/09/0238252 ). The same concept but in decomposition. If every human profile could be cut down to, say, 50 polygons and we just stored their position and orientation, a relatively simple record could be kept of each person.
Still too many thoughts for me to be getting it right though.
Me failed English...
FreeBSD over Linux. If my comments seem odd, this may explain...
Friends, I never tire of repeating: Google is the NSA.
Once a clever guy in the NSA (they have lots of them) pondered... and found a way to get people giving them willingly all sorts of info. On top of that, it's an operation financing itself!
Hear hear. I value "radical openness," and it seems to be a minority opinion sometimes. I don't mind if strangers can find me; I find it interesting to see a picture of an author or a friend of a friend that I've heard lots about. And I don't mind if people do the same for me. Mod parent up!
http://www.idmt.fraunhofer.de/eng/research_topics/photoid.htm
In fact human faces can be compressed very effectively. The top 20 features from eigenfaces are more than enough for recognition, so forget polygons.
As someone who has also worked in the computer vision field, I agree completely. However, I have found that it is sadly pointless to discuss out how this stuff really works on slashdot, as the only comments that will be modded up have to do with tin-foil hat like predictions. You can point out that a national database of faces is unrealistic, as techniques such as Principle Components Analysis aren't accurate on massive data sets, and your comment will squander at 1. Don't even bother pointing out all of the inaccuracies in the article, such as how face detection commonly works. Most people on this site think that this all works like it is shown in the movies.
The exact opposite is the case with PCA, all too often it is applied to tiny datasets, where it cannot really be useful in extracting meaningful relationships from multivariate data.
The larger the sample size, and the larger the number of variables for each sample, then the more likely it is that PCA will actually be useful to you in reducing the dimensionality of the dataset to enable human observation of reliable and repeatable relationships.
The smaller the sample size and the fewer data points per sample, the more likely it is that you will be able to plainly see any relationships in the data, and any application of PCA is simply so you can say, look, I'm cool, I used Principle Components Analysis!