|
|
Face Detection Hi everybody, i already read many treads with experience in face detection, but i didn't find what im searching for.. i would like to make a face recognition for a project.. QC must recognize the face and make something with this recognition.. the most close experience is make by DUST ( http://kineme.net/composition/dust/facedetector ) but when i tried to open it this error appear: could someone give me an help? thank you very much guys!
|
You need to be running with a version of QC with support for the Face Detector (using FaceCoreLight) patch. If you are running SL, or earlier you won't find it. Lion and upwards - you'll be good to go, [or should be]. That said, it is still a technology in development.
Check out the CVTools patch that's available here. There's already an included face "haar" detector file, if I'm not mistaken, in the corresponding sample composition download zip. There are some other "face" haar trackers on the web, and haar tracker files for eye, nose, mouth, etc., that can be loaded by CVTools.
@cybero thanks, i suspect before that wasn't for SL but now i've got a confirm.
@gtoledo3 i will follow your advise, and ill give it a try. may i ask your help if i'll be in trouble?
Sure, feel free. Hope it works out.
Hi here i am :) in attachment there's a composition that i use for my test.. its the haar sample modified for what i supposed to do, wearable eyeglasses. i cant manage that eyeglasses move exactly with the head..what type of mistake i've done?
thank in advance to everybody can help me.
Can you include whatever "haar" pattern file you are using? It would be hard to check the composition without that (I haven't looked at the qtz yet).
ups, sorry.
here it is.
thanks.
Well, I looked at the composition. To me, it seemed like the Y offset was off, and can just be 0.
That said, take a look at the stock haar tracker composition from kineme again - the one with the face targets.
What you'll notice is that the image is iterated and also sized so it gets bigger or smaller depending on some factor (I can't remember where it gets that info). I know that tweaking that parameter is important to get the left/right, up/down "scale" of motion in line with what you expect as the haar analysis determines that you're closer or further away from the camera.
I'm not sure if putting everything in that Render In Image threw off any of the scale of motion or not... the sprite should be sized the same as the source image or rendering destination, probably not 2 and 1.5 though that's close. I don't think that's what throws off the x-axis here though.
Looking again...it seems really weird to me that the haar is then going through an optical flow analysis, that seems like it would throw it off. Your thresholds may be reversed as well. Take a look at the kineme sample composition and pop your glasses image into that, and work backwards from there.
the patch your using is from a really old hack when the haar patch wasn't working properly. that has been fixed for sometime. so there isn't really a need to run the haar into optical flow anymore. however it does still work. simply adjusting an x and y offset sticks the glasses to my eyes. you should look at the new open cv example comps like george mentioned to get a example of how to use width and height etc... here these offsets work for me.