Face Recognition On Mobile Phones

gpvillamil writes: "This article describes a collaboration between Motorola, Visionics and Wirehound to build in an automatic mug shot recognition capability into mobile phones. Particularly interesting is how the phones will scan all faces in the field of view, and indicate matches by an instant short message."

Comments

Re:Right to privacy?!?

[swillden]

Since when do you have an expectation of privacy sitting in a cafe?

You have none. However, the poster does have a real point, just bad terminology.

Obviously you have no reasonable expectation of privacy in a public place: Anyone who happens to look (or point a camera) in your direction can see whatever it is that you're doing. You do have an expectation of some level of anonymity in many public places, however, and that brings with it a sort of privacy. Anyone can see what you're doing, but no one knows who you are.

But this expectation of anonymity is a very relative thing, and it's also a very *new* thing if you look at all of human history. Even today, residents of very small towns (say, less than a thousand people) have no expectation of anonymity in a public place. Anonymity in public is a phenomenon that arose first in large cities and more recently in very mobile populations.

Is public anonymity a good thing? Maybe. A necessary thing? Clearly not (I think privacy *is* a necessary thing).

One way to look at this is just as a continuation of the development of the "global village". That term is usually used to indicate the ease with which people throughout the world can communicate and trade, but one aspect of a real "village" is that no one in it is a complete stranger. Having lived in a very small town, I can tell you that everyone knowing everyone else has both advantages and disadvantages. In a small town, you know who you can trust and who you can't (and there are typically very few of the latter, which is why many people in small towns don't lock their doors). OTOH, it really sucks for the person who does one stupid thing one time and puts themselves into the "can't trust" category.

All of this presumes that the mobile face recognition technology (or any face recognitiion technology, for that matter) really works, which at present it does not. Current face recognition technology does quite a good job at comparing an image and a template and deciding "is this probably the same person?" Good in the sense that it's fairly difficult to pass yourself off as someone else. However, when you change the problem to matching an image against a database of templates and deciding "Which person, if any, is this one?" then current FR tech is horribly bad. The birthday paradox makes this a really hard problem to scale.

Given the fact that even humans, with their relatively small databases of a few thousand faces and highly developed recognition abilities, occasionally make mis-identifications, it's by no means certain that computers will ever be able to do this well on a large scale.

Re:Right to privacy?!?

[swillden]

It occurs to me that not everyone knows what the birthday paradox is:

For those who aren't familiar with it, the birthday paradox comes from the old party game question "Are there two people in this room with the same birthday?" The naive, commonsense answer would consider that there 366 possible birthdays in a year and expect that out of, say, a group of 30 people there's only a roughly 10% probability that two have the same birthday. That would be a good estimate of the probability of one person in the room having a particular birthday, but the problem is that in the room there are 870 (30*29) different "pairs" of birthdays, and the odds one pairing has the same birthday is quite high (I'm too lazy to calculate it, but it's around 70%).

In the case of face recognition and other identification technologies, it's like having 300 million people in the room and the problem is how to extract enough information from a face to create a relatively "unique" facial signature, because if there's even a small probability that signatures of two different faces will be the same, you will get *lots* of false pairings. Suppose that there is a 10e-6 probability that a random image will match a random template. In a nation of 300 million that means there will be, on average, 300 other people that, to the computer, look just like you. Worse, if you have a database of, say 1,000 bad guys, and you have an airport with 100,000 people passing through it daily, you'll get an average of 100 false positives *per day*.

And current FR tech is a long way from the 10e-6 false positive level. More like 10e-3 (meaing in the airport scenario that for every passenger passing through the terminal, the computer will find one possible match in the bad guy database. On average, of course).