Thursday, 20 August 2009

Begs the question

Chris's comment made me realize that there was another option in the Mohamud case.

Why couldn't they use Facial Recognition Software when the photo ID in the passport was disputed?

Tests have shown that computer analysis is more accurate than human (and would have corroborated the experts in this case.)


Christopher Parsons said...

Joseph Ferenbok at UoT is doing really good work on the face, with focuses on facial recognition. Based on conversations with him, I'd suggest that many of the 'reports' about the effectiveness of FRT are substandard or at least wanting. It's still a technology in its infancy, and while 3d imaging is better, it's still a ways off. Of course, 2d analysis is typically plagued with problems.

Catelli said...


Many laptops now have cameras built into them and I know companies are pushing facial recognition as a means of authentication. But I haven't played with it.

I found reports from just 5 years ago that criticized the accuracy of FR software, but recent reports suggest improvement is being made.

I guess that uncertainty is what would prevent adoption.

Oh well, it was a thought.

Christopher Parsons said...

The FRT systems on laptops with cameras are generally junk. I've seen various situations where they are easily defeated - most recently there was a video blogger who showed how their multi-thousand dollar, brand new, Thinkpad facial recognition was defeated using a printed picture of someone of the same general race (Asian, in this case). The picture was grainy and only offered the most general of details on the face, but the FRT system saw a match and let the guy into the system. Many of the FRT technology, at the moment, is deeply sensitive to white faces, and highly inaccurate when it comes to non-white faces.

2d analysis is really hard to pull off well. This isn't to say it won't improve - I think that the massive insertion of images into government databases that meet standards for FRT research will be massively helpful in improving FRT systems. 3d is better, because it does map the curves of the face, but more costly to get data and apply algorythims (as I understand it).

There are certainly improvements, but we're still talking about, on average, a 60-70% accuracy at absolute best, under ideal circumstances. Examples of it being used right now are in gambling establishments in Ontario casinos, where problem gamblers can submit their images. As a person approaches the entrance, if it is suspected that they are one of the problem gamblers their ID is screened against the profile that is automatically called up. Their system is OK, because it tries to eliminate false positive by physical document identification as well, but that the two-positive system is implemented demonstrates that even where (literally) billions have been spent in FRT that it still has a ways to go.

Catelli said...


Bubble deflated!