Cycling 74 » Topic: good face tracking ?

good face tracking ?

Dieter_Laser — Fri, 06 Nov 2009 00:17:33 +0000

hello,

im not happy with the tracking algorithms of the cv-library for max. you cant track the profile of a face with this.

does somebody knows good librarys or programms i could use togehter with max, that track almost any faceposition for more than one face ?

that would be great.

greetz

Re: good face tracking ?

yair reshef — Wed, 18 Nov 2009 17:56:12 +0000

faceAPI (windows only) offer a great face tracker for non commercial use (limited to one face).
it does take a moment to “lock on” but will track faces in odd angles. i haven’t seen better.
http://www.youtube.com/watch?v=7P7I7DOJJB0 (warning stupid audio)

there’s also the crazy russains but they use opencv and dont share

Re: good face tracking ?

readywater — Fri, 20 Nov 2009 16:50:34 +0000

I haven’t played with the cv.jit library too much, but see if it’s possible to change the haar-like cascade. Wikipedia is a good start to learning about computer vision, but the o’rielly media book “OpenCV” is great: highly recommend.

Re: good face tracking ?

i.te — Sun, 22 Nov 2009 19:57:08 +0000

eyesweb http://www.infomus.org/EywMain.html

eyecon http://www.frieder-weiss.de/eyecon/download.html

vvvv with open source library for contour tracking.

Re: good face tracking ?

ygreq — Wed, 23 Jun 2010 19:46:32 +0000

I found out that if I light up the scene pretty well (soft and bright), I have no problem detecting faces with cv.jit.faces, profiles and so on!

Re: good face tracking ?

artm — Wed, 23 Jun 2010 20:38:30 +0000

yes, you can change the cascade in [cv.jit.faces], send a message (read “cascade-filename.xml”), a file “haarcascade_profileface.xml” comes OpenCV among others.

also combining haar cascade detection with camshift tracking works well (detect a face, initialize camshift to track colors in the face region then use camshift to track the face region, which will continue to work even when haar feature detection doesn’t (actually you can/should switch off [cv.jit.faces] until camshift loses its target).

I don’t have an example in jitter though, I only did this in plan C ( http://www.youtube.com/user/femistofel#p/a/u/1/n18cGiZ8DGQ )

Re: good face tracking ?

Jean-Marc Pelletier — Mon, 05 Jul 2010 14:05:29 +0000

Hi,

The OpenCV 2.0 release has a decent profile face detection cascade, which I’ve used successfully in the past.

I put it up here: http://jmpelletier.com/data/haarFaceProfileCascade1.xml.zip (This is originally part of OpenCV, so please read the relevant part of the license.)

You’ll have to flip your image, though, as this only recognizes faces pointing in a single direction. All in all, you’ll need three cv.jit.faces objects.

cv.jit.faces uses OpenCV functionality pretty much as-is, so I take no credit for its performance. It is, however, based on the most widely used algorithm around. There may be better methods around, but this one has been proving itself for a few years now and it’s still pretty close to state of the art. However, it is fairly dependent on the quality of the cascades you use, although the ones that ship with OpenCV and which I redistribute are pretty decent.

As artm mentioned, cv.jit.faces does face _detection_, which is different from face tracking. You can use trackers such as cv.jit.track and cv.jit.shift in tandem to get much better performance, but it’s more work of course.

Jean-Marc