[varelse]

Given the 15.8% success rate of the cat detector on large data sets: http://arxiv.org/pdf/1112.6209.pdf&embedded=true,

My chief worry is overconfidence from hopelessly technologically ignorant bureaucrats combined with a slew of false negatives and false positives when applied to human faces on general surveillance feeds.

In contrast, driver's license, ID, and arrest photos are at least severely constrained input filters wherein a single individual is placed against relatively uncomplicated background pixels and the face is in roughly the same orientation.

[nathan_long]

I don't know about the cat detector, but Picasa seems able to identify people in photos pretty well, despite varying positions and lighting conditions.

It creeps me out.

Granted, they are working with the much smaller set of people I know rather than the whole populace. But technology only gets better.

[jerf]

"But technology only gets better."

Technology is bounded by mathematical constraints, though. The more you constrain the number of faces you are scanning and you are looking for, the easier the task is. But when you're trying to pick out any of perhaps thousands of persons of interest across an input set that is, say, in the millions, the more impossible it becomes.

If it helps, imagine that you've equipped a person with all the people they're looking for, and then set them up in front of the cameras. Even if you imagine a superhuman who somehow has time to examine all of the millions of inputs, even the human is going to have a huge number of false positives and negatives. The problem itself is fundamentally, mathematically hard. Any surveillance technique based on the idea that computer vision is better than human vision at this sort of thing is going to fail, hard.

And that's before the population starts taking active countermeasures against the face metrics. (As usual, despite the abundant evidence we live in a dynamic, reactive universe, most humans persist in functioning in a static model, where the Bad Guys will just passively sit and let their faces be scanned and never react to that.)

[varelse]

Just like there was a relatively cheap countermeasure to every pipe dream revision of the SDI, such countermeasures are already in place for facial recognition:

http://news.discovery.com/tech/gear-and-gadgets/glasses-foil...

But at least we'll fall victim to a higher, rarified caliber of terrorist rather than that awful riffraff we're dealing with now.

[jerf]

Those aren't even the interesting countermeasures. Good work with makeup can significantly change the structure of your face, completely reversibly, far more than is necessary to fool these systems running across the full set of human beings. You don't have to get the face recognizer to not recognize a face at all, you just have to make it a wrong face.

[cbhl]

Identifying people's faces, especially caucasians, is relatively considered straightforward nowadays. Google spent a lot of time working on the problem so they could obscure faces in Google Street View. You start by looking for the eyes, which are dark spots surrounded by white. Then, you go from the eyes to the nose.

Identifying which person a face belongs to is a classification problem, which gets difficult the larger the set you're comparing to. Facebook has some tech that is interesting here, but it also only checks your friends and some level of false positives is acceptable. I imagine it that what algorithms they have aren't quite good enough to, say, identify your face out of an entire city's worth of driver's license photos. (Yet, which is what this article alludes to.)