[ml] Hi

[Josh Myer]

On Mon, Jul 18, 2011 at 1:50 PM, Mike Schachter <mike at mindmech.com> wrote:

> Hey Alexi,
>
> Sounds pretty cool. Have you checked out the mobile robotics
> toolkit? I haven't used it but MC Hawking (the semi-autonomous
> wheelchair robot at Noisebridge) uses it:
>
> http://www.mrpt.org/
>
> Idris, that sounds like a pretty cool project! If you're doing image
> processing, you might want to look into OpenCV:
>
> http://opencv.willowgarage.com/wiki/
>
> They have some built in machine learning algorithms for
> classification (i.e. here's an image, does it contain a tennis
> court or not?)
>
> http://opencv.willowgarage.com/documentation/cpp/ml__machine_learning.html
>
> You would need to feed it examples that are already trained
> though. I wonder if you could use it's 2D feature extraction stuff
> to algorithmically decide if it's a tennis court or not?
>
>
Unsolicited kibitzing from a grumpy old practitioner:

I'd suggest looking at SIFT, though it may have problems with the different
colors in the courts (I'm not entirely certain how it's identifying
features).  Pushing things through a mapping (desaturate and/or edge detect)
will take care of that.  It's probably helpful to reduce the resolution,
too, to eliminate confounding things like fence shadows.

http://blogs.oregonstate.edu/hess/code/sift/

As far as algorithmic detection of "is tennis court?": you shouldn't have
too hard of a time manually finding tennis courts in a handful of
cities/towns to build a training set.  You'll want to find a bunch of
confusing not-tennis-courts for your training set, as well: roofs, fields,
landing strips, etc.

Personally, I'd start with an image of a tennis court, and make sure I can
identify it when rotated, scaled, and color-shifted.  That's probably a fair
bit of work, but is prerequisite to most other techniques.

Always remember: 95% of ML is scut work, 5% is sexy algorithmic fun, and 50%
is refactoring your code to account for the stupid ugly data gotchas you
missed in the original scope.

Happy hacking!
--
/jbm
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.noisebridge.net/pipermail/ml/attachments/20110718/bd8f941b/attachment-0003.html>