Yep, FB has an enormous corpus of accurate training data due to its millions of users performing tagging on hundreds or thousands of photos each. I would venture to say that they're the only ones that could do this/enable this to be done.

I've done this kind of research 4 years ago based on general Internet image. That was a failure, but because of the rising of Facebook, it might work again. Someone already work on this for sometime: face.com.

Ive talked with one of the guys from face.com . They have put a lot of effort into making their system very fast and scalable.

He told me that they built it from scratch and did not use OpenCV. Some parts were even written in assembly.

Im pretty sure that this is the only web application that I use that has had major parts of it written in assembly.

I agree that for some specific tasks OpenCV is not particularly useful (I have to write detector and trainer from scratch even there is opencv haartraining. The OpenCV crowd are most interested in applying it to real-time video processing and put little effort on Internet scale problem).

But for assembly optimization it is really not as useful as it used to be. Scalability is more about how to get sub-linear time complexity and efficient communication pattern. Nowadays c compiler can get fairly good assembly code and for low level optimization, human cannot compete with machine (how many people knows the particular cache line alignment trick on old core i7?). Multimedia instructions (SSE/MMX/3D Now) are useful but most of them can be done by function call instead of hand-crafted assembly.

And the GPS info could narrow down the search in most cases.

> of course, efficiently searching thru a corpus of millions of faces

There was research cited here a while ago on the high accuracy obtained by constraining the search space to within one's social graph (hundreds vs millions). While it might not identify the stranger at the bar, it might identify their companions and so on. Six degrees of separation in real-time.

Add in security camera networks and you get all sorts of geographic applications without the need for always-on GPS!

For me this won't be very useful until someone develops the more crucial ass recognition technology.