Skin Detector is an attempt at writing a vector accelerated (both altivec and SSE) algorithm to detect human faces based on well known properties of the human skin.
It is based on the assumption that
human skin fits a color, hue and saturation envelope which varies
little across multiple races. While there are many other approaches
to this subject, such as template matching or feature detection,
the sole purpose of this project was to implement the detection
algorithm in vector accelerated code. I hope to have the time to
improve it with features such as eye, mouth and nose detection
inside a human-skin candidate area to determine whether a human
face is present. This, combined with a face vector template
database, could allow subject identification across a variety of
orientations.
Can i check it
out?
Sure, you can download the sandbox here, but be advised that
several light balancing and camera gain compensations are
hardcoded, so you may not be able to see it's magic.
The next sandbox will have scrollers to adjust what should be
adjusted and auto compenations for what should not be adjusted.
Untill that happens, you'll just have to wait.
See it in
action
Left: PowerBook Ti 1GHz and QuickCam 4000 / Right: iMac G5 20" and
built-in iSight
(First tech preview - download version - uses all masks except for
HE)
1) Click here -
Download tech preview. Manually adjust threshold values to match
skin. Auto-calibration was added after this compile.
2) Click here -
First attempt at locking my skin tone under artificial light.
3) Click here
- Using offscreen blur, erode and pixellate masks to improve
tracking. Shaved hair and face :)
What about the
source
The source is a mess. It's not fully documented nor properly
segmented into classes and selectors. Most of the altivec stuff is
NOT packed in a separate library of utilities (such as vector
accelerated arc-tangents, base-10 and natural logarithms). I do
plan on releasing the source but only when i get it cleaned
up.
What technologies
does it use?
Cocoa, Objective-C and plain old C. It used the Accelerate
framework which provides a common API to both SSE (Intel) and
Altivec (PowerPC) SIMD units. I haven't figured out if all of the
framework's functions are available in both units or if some of
them are exclusive to either unit, because i haven't had the time
to read the docs and i don't have an Intel machine with a video
grabber and a legit OS X to test if my code runs on SSE.
Beware: This application does NOT CHECK for a vector unit when
lauched. It assumes there is one.
UPDATE: Optimized algorithm to run on 70% CPU load on PPC G4@1Ghz.
No need for Vector unit for now!