General Misc > Robot Videos

Facial Recognition (Lecture)

(1/2) > >>

Admin:
Im currently working on the third tutorial in my computer vision tutorial series, so what better time than to offer a supplementary?

A lecture on facial recognition:
http://mitworld.mit.edu/video/154/
(1 hour)

"He theorizes that facial perception is a holistic process: we broadly take in the relationship, for instance, of eyes, nose and mouth. He tested this hypothesis by creating a computer program that could similarly grasp facial structure, and the program was able to “see” a face within a larger picture. In his Hirschfeld Project, Sinha is trying to distill the caricaturists’ understanding about the important landmarks of a face."

Key Points
- Vision is our most important sensor.

- Facial recognition will greatly enhance AI on mobile robots.

- "Parallel and Hierarchical Model of Vision Processing" - dividing up different vision tasks as entirely seperate

- edge detector neurons

- action potentials, signals from neurons, tuned to seeing different things

- object recognition, very fundamental problem in neuroscience

- over 1200 studies done in facial recognition in the last two decades

- image quality is important for facial recognition

- spatial configuration important (location of eyes, nose, mouth, etc.)

- facial recognition reaction time increases with lower quality images

- removing facial image degradation based on training techniques

- which aspects of facial recognition are important, and which are not?

- characterture artists 'understand facial recognition intuitively'

- facial recognition and object recognition are very similar tasks

- how does the mind collect and store object recognition information?

dunk:
for anyone who is interested in Facial Recognition (or any other software based image processing) you should definitely look at this library:
http://www.intel.com/technology/computing/opencv/index.htm

they have a sample application that will pick faces in realtime out of a live video feed.
smile for the camera!

dunk.

JesseWelling:
 man it's just like an aim bot in counter strike...  :D

Cognaut:
Interesting lecture.  It stimulated my process.

One thing that I wonder is that, in today's technology, "blur" is most commonly recognized as a product of insufficient resolution.  We recognize the pixelation.  It's our perception though, that our eyes are not pixelated.

In the continuous world, we encounter types of bluring that have no recognizable graduations.  Like when staring at the computer screen too long, when the windows are fogged up or when spear fishing.

We have the ability to detect and approximate visual distortions in order to see through them.  I think the key to this is that we learn to simulate the distortion and use it for comparitive reference when trying to infer what's beyond it.  Not mentioned in the lecture is the role of stereoscopy and motion in the process.

Imagine a smart camera that produced 3D vector files instead of jpgs.  The files could be rendered with any resolution desired and viewed in stereo.

JonHylands:

--- Quote from: Cognaut on December 11, 2006, 01:16:26 PM ---Imagine a smart camera that produced 3D vector files instead of jpgs.  The files could be rendered with any resolution desired and viewed in stereo.

--- End quote ---

You've just described the key to my vision system design - the real trick is figuring out a way to do this in real time (which I haven't lyet)...

- Jon

Navigation

[0] Message Index

[#] Next page