Making The World Safe For Science - One Supermodel At A Time
Face Recognition, Emotions And Why I Will Be The Next Jackie Chan
Dynamic face recognition may be one of the last frontiers for quantifying why some of our simplest brain functions have difficulty being matched by computers.
As a kid, when computers were more basic, it was easy to see the differences between my brain and the power of a computer. When I played baseball I knew the instant the ball was hit whether I had to run back or forward. This was a spectacular amount of mathematics done in an instant with the only programming being my amount of practice. A computer has to be able to define a parabola before it can tell me which way to run - by the time it could tell me where to run, the ball will have landed. The only reason I bothered to show up for trigonometry was to figure out a methodology that could help me write a computer program to catch a baseball without having the computer know where the ball was going to be hit in advance.
Facial recognition today faces similar obstacles. If you look at your favorite scientist's irritated face while his supermodel storms out because he did not tell her she was beautiful for the requisite hundredth time that day, it looks a certain way. If he then looks happy because he finally got some peace and quiet, it looks another way. To you, it is easily the same person but a computer has a very difficult time with that operation.
A three-year old child can look at a real chicken and then a cartoon chicken and say, "That's a chicken." A computer has a very difficult time with that also. Snd good luck getting a computer to tell you who is more attractive, that girl from "Veronica Mars" or the live action model of Lara Craft, Tomb Raider:
In the late 1970s research began to kick into gear, mostly because of vision research. Scientists wanted to determine how much information was essential for visual recognition. Similar to hearing, we can determine frequencies of interest for vision. In this case, the frequencies are spatial and the research was done by AP Ginsburg.* He used bandpass filters to determine which spatial frequencies carried the most information and therefore could be isolated to make recognition easier. He narrowed down the important range of frequencies but, also important, he discovered that different spatial-frequency sets supplied the information for different perceptual and cognitive functions - like emotions.
It's only now that home computers will contain the horsepower to do this kind of number crunching. The field of identity recognition is easily covered today but emotional recognition - the kind of thing that allows a computer to recognize me regardless of my supermodel irritation level - still has some way to go.
I've tried to find some legitimate modern research that makes this big improvement but so far it just involves brute-force techniques and more computing power to identify and match grid-by-grid; hardly an elegant solution.
Yet while emotional variance recognition is still lacking ordinary face recognition has grown tremendously. I went to a site called My Heritage and tried their tool. The demo version allows you to input your picture and tells you what celebrities you look like. Naturally, I had to give that a try. So I threw in my picture and it comes back with the four people in the graphic.
MyHeritage.com said I am a combination of Trent Rezner, Scott Bakula and Jackie Chan. I wasn't sure how that Colin Firth guy fit in there but I remembered he nailed Keira Knightley in some movie or another, so it makes sense now.
They think I am a rock star, a quantum physicist and the greatest martial arts action hero of all time? This software is so accurate it is almost spooky.
*Ginsburg A.P. "Spatial filtering and vision: implications for normal and abnormal vision." In Clinical Applications of Visual Psychophysics. LM Proenz, JM Enoch, A Jampolsky (ed) Cambridge University Press, 1981, pp 70-106. I can't find an online link.
posted by Buckaroo at 8:42 AM