Hi David,
what I am aiming at is not a totally equivalent of OpenCv running with
the full Davinci power - which is far too ambitious from my point of
view (standard OpenCV code will never compile on a DSP). As I wrote in
post and the follow up
http://linux.omap.com/pipermail/davinci-linux-open-source/2008-May/006685.html
I am basically rewriting only the most important cxcore.h-functions, and
only those a DSP is much faster with. Next to that, I have written a
nice framework around that so one can use DSP functions with the normal
OpenCV datastructures. Especially the dynamic allocation works now, but
it's kind of a hack.
The most interesting functions will surely be YUV-conversion,
convolution and matrix multiplication. These can speed up some image
recognition tasks.
At the moment, I am in fact planning some face detection task for my
thesis presentation. I got the normal OpenCV face detection demo easily
running on the ARM core, but only with 3 FPS. Perhaps I will be able to
speed it up to some extent when YUV-conversion and cvIntegral will be
done on the DSP to roughly 5 FPS. The problem is that face detection
algorithms normally use very complex datastructures for the sample data.
And of course, the algorithms are complex themselves. If you have ever
programmed a DSP you will not be trying to interpret a Haar features
cascade database within a DSP program. So only some sub functions of it
can be ported to the DSP in a given amount of time.
Of course, other object recognition algorithms like i.e. SIFT could be
good candidates for parallelization.
-- Andre
David Greco schrieb:
what news about you porting?
did you finish it?
_______________________________________________
Davinci-linux-open-source mailing list
[email protected]
http://linux.davincidsp.com/mailman/listinfo/davinci-linux-open-source
