On computer face recognition:

geeman at best.com geeman at best.com
Fri Apr 12 06:01:36 PDT 1996


There was a piece, I _think_ in Scientific American, tho it might have been
an AI journal, on face recognition by use of neural nets together with what
were called "eigenface" images: 
These eigenfaces each have specific characteristics, which when combined
together can closely approximate a specific face image.  The target face was
analyzed in terms of closeness-of-match to a small set of eigenfaces, on the
order of 5 to 8, I think.  Results of course were promising (else why write
about it, eh?) if not excellent.

At 12:14 PM 4/11/96 -0700, you wrote:
>At 10:02 PM 4/10/96 -0700, you wrote:
>>
>>On computer face recognition:
>>
>>
>>>> Shaving probably will not be a problem, but holding your head at a
>>>> slightly different angle...  will screw up the system totally,
>>>> unless the system has radically improved since the last time I read
>>>> up on it.
>>
>>At 11:45 AM 4/9/96 -0500, K00l Secrets wrote:
>>> Well, the systems I have seen are quite good at finding people's eyes.
>>> Scaling (for distance), and rotation (for the angle of your head)
>>> therefore don't really confuse the system once it has your eyes.
>>
>>Finding the eyes can only control for rotations in the plane of 
>>the image, when you tilt your head to one side.  They cannot 
>>handle the much more common case of 3D rotations, where you 
>>look slightly to the right or slightly to the left of camera.
>>Facial expressions also throw them badly.
>>
>
>Take a peak at http://www.neci.nj.nec.com/homepages/lawrence/papers. One of
>Lawences papers is on using Neural networks to recognize faces. Methinks
>that the state of the art is advancing rapidly and such problems as not
>looking at the camera or changing your expression are rapidly being overcome.
>
>
>=====================================
>dwl at hnc.com
>
>Zippity do da, zippity ah, my oh my what a wonderful day.
>Ya right, and hear I am without time to finish a cup of coffee.
>
>







More information about the cypherpunks-legacy mailing list