A Face in the Crowd

One of our recent class assignments was to create a blog post based on an original research article from Nature, Science, or if all else failed, the National Academy of Sciences. In my ignorance, I assumed this would be easy. Everyone cares as much about my topic as I do, right? Wrong!

So I expanded my searching to include the supplementary areas of national security, areas that can in turn create an even more well-rounded program. One interesting article I came across was 100% Accuracy in Automatic Face Recognition. (This link leads to another posting of the paper, but you can also view it through Science if you have access to the magazine.)

Though the article was published in 2008, it's still incredibly relevant to our national security efforts as more cameras and software are included in major public places. This article is even more pertinent now with the advances of technology, and in turn photographs/videos, in our everyday lives -- Facebook, anyone?

The important piece of the article is that on a one-by-one basis, a positive confirmation of photo identity averaged 54%. This means that when the researchers compared the photo they wanted to confirm against the images in the database, it was only correct half the time! The primary reason the software wasn't able to correctly match the photos was because of variability -- lighting, poses, age, quality of the photograph, etc.

To help improve this "hit rate", the researchers created an average image from the databases. In the photo below, you can see how various photos can be combined to create a single image with many variables included. Comparing the averaged photo against each of the photos available in the database created a 100% confirmation rate! However, the researchers thought this seemed unfair since the averaged photo contained the photos that were originally confirmed. It's like doing well on a test when you've had a chance to look over the answers in advance.

To resolve this, the researchers created another set of averaged photos. These were comprised of only the photos that weren't confirmed originally. So, all the "failed" photos made up a new averaged photo. This time, the software was able to confirm a "hit" with 80% accuracy. Not bad, considering the new photo was made up entirely of photos that failed the first time around.

So why would this be important now? Imagine someone coming through an major public place, and a security camera catches a glimpse of them. Unfortunately, the picture the security camera takes isn't great -- grainy, out of focus, dark, whatever. These types of software improvements might be able to confirm the person is someone of suspicion, and flag a human to confirm. Every tool law enforcement has at their disposal could be important, depending on the current situation. As long as we're using many tools, approaches, techniques, etc, we're stronger than we are when dependent on only a few.