The VXL Project

> -----Original Message-----
> From:  Miguel A Figueroa-Villanneva

> 1. I would like to know if these functions are well developed or are =
they
> still at an early experimental stage? Also, are there any other =
packages
> that I might have overlooked that will help me solve the problem?

ipts contains some relatively simple implementations of published =
algorithms. It isn't under active research development, because the =
research group at Manchester isn't doing deep research into new interest =
point detectors, but how existing published detectors can be used for =
appearance modelling, object location, detection, etc. I believe that =
the code in oxl has been stable for some time.


> 2. Is there any interest on your (vxl developers) part for the view
> morphing functionality to be contributed to vxl? If so, is there any
> documentation or guidelines on how to do this (i.e. who do I submit =
code
> to, and what programming guidelines besides those in the VXL Book =
should I
> follow)?

We are always interested in new code. The book is the primary reference, =
but a few other things can be found in the vxl/core/doc directory, =
including records of style decisions taken at the Zurich meeting a few =
years ago. If you are providing an add-on library, there are few =
aditional restrictions. It should compile, preferably without warnings, =
on the platforms on our dashboard. It should contain lots of API level =
documentation. Changes/additions to core libraries are treated a bit =
more strictly, and often subject to post-submission ad-hoc, but =
friendly, peer review. To actually submit stuff, get yourself a =
SourceForge username, and we will add you to the commit list. Finally, =
subscribe to the vxl-maintainers list, where you should send any =
questions about submissions, etc.

> 3. Since my application needs to run in real-time, I would like to ask =
if
> anybody has any feedback on preformance comparisons of vxl and intel
> opencv. I think vxl is a more complete package, however, I'm working =
on a
> performance critical application. Any objective feedback will be
> appreciated.

OpenCV probably has more "published computer vision" algorithms in it, =
whereas VXL was designed from the ground up as a series of complete core =
libraries. Given the effort Intel have put into optimising those =
algorithms, I would imagine that most comparable complex algorithms =
probably run a bit faster in OpenCV. I would be surprised if the =
difference is more than a small constant factor. One area that has been =
measured was comparing Intel's precompiled implementation of BLAS (the =
MKL math library I think is used by OpenCV) with gcc -O3 and icc -O3 =
compilation of VXL's math library vnl. For vector-vector and =
vector-matrix operations VXL was generally faster, with no consistent =
difference between gcc and icc. As with all computation, the best way to =
get speed is to write your program as simply and algorithmically =
efficiently as possible, compile it with the best compiler you have. =
Then profile it, and spend a bit of effort fiddling with the small =
number of inner loops where your program spends 90% of it's time. This =
approach has resulted in several important bits of VXL being optimised =
to the point (on x86 anyway) where the inner loop's compilation looks =
identical to what a hand written assembler version would look like.=20


Ian Scott
University of Manchester