Slashdot Mirror
US Intelligence Wants To Radically Advance Facial Recognition Software
coondoggie writes "Identifying people from video streams or boatloads of images can be a daunting task for humans and computers. But a 4-year development program set to start in April 2014 known as Janus aims to develop software and algorithms that erase those problems and could radically alter the facial recognition world as we know it. Funded by the Office of the Director of National Intelligence's 'high-risk, high-payoff research' group, Intelligence Advanced Research Projects Activity (IARPA) Janus 'seeks to improve face recognition performance using representations developed from real-world video and images instead of from calibrated and constrained collections.'"
For similar reasons as described in https://www.schneier.com/blog/archives/2012/05/criminal_intent.html it will not be usefull.
I've worked with current facial recognition systems and they're absolutely junk. They can match mug shots with perfect lighting but that's about all. It's a very long way to being able to pick people out of some crappy live video stream. Mind, I worked with whatever's publicly available; maybe the various big brother agencies have better stuff; i wouldn't bet on it though.
A while ago I did a little research in computer vision. From the summary it seems like nothing more than moving a project from an academic project to a real world project.
In the academic world it is perfectly acceptable to use carefully selected or crafted inputs (facial images in this case) to develop and evaluate your algorithms. You may have separate date sets for development and evaluation, however careful selection or crafting is OK to simplify the project and avoid issues/variables outside of the project's scope. In your particular mugshot example this would be using images of good resolution and good/predictable lighting. Dealing with low resolution and bad lighting would be an issue left to the next thesis or research grant or for commercialization.
Working with mugshots may be a fluke, the inputs happen to be carefully crafted like one might do in academic research. So it was relatively simple to transition to this niche real world application.
Moving to a general real world solution using images and video of questionable quality is an enormous jump in the level of difficulty. Perhaps too difficult. It may not be possible to recognize an individual. It may only be possible to offer a somewhat generalized characterization that a person my fit into. At least with the haphazardly placed cameras typically found on the streets and in shops today. Some places use very good and carefully positioned cameras to get decent images for automated facial recognition. For example Las Vegas casinos.