[PD] openCV face detection performance Was: make me

Mathieu Bouchard matju at artengine.ca
Thu Aug 19 18:42:23 CEST 2010

On Thu, 19 Aug 2010, Max wrote:

> While I was checking out a Canon EOS 5D automatic face detection and 
> focus, I wondered why this camera can do so effortless and the OpenCV on 
> a Mac 8 core machine uses one Processor with 250% for that. When 
> changing to greyscale colorspace and reducing the resolution to a 
> quarter of PAL one get under 100% but it still still seems like a lot. 
> Any Ideas why that could be?

There are several reasons, one of which being that specialised processors 
can still be much faster than general processors (just look at what video 
cards can do...) but it's also very possible that they downscale the image 
by a factor of 2*2 (4 times faster) or 4*4 (16 times faster) before 
starting to apply the face detection.

Note also that converting from GEM's YUV to RGB upscales the U and V 
channel by a factor of 2 horizontally (no vertical scaling), which means 
that the RGB data is a bit redundant. That's the "422" mode. For cameras 
in YUV "420" mode, there's also vertical scaling, therefore, in my camera 
code, I have added an option to convert directly from YUV420 to a twice 
narrower, twice shorter RGB picture, for the case where the chroma 
resolution is absolutely essential. It reduces a lot of waste.

  _ _ __ ___ _____ ________ _____________ _____________________ ...
| Mathieu Bouchard, Montréal, Québec. téléphone: +1.514.383.3801

More information about the Pd-list mailing list