Re: Looking for a segmentation algorithm

On Sun, 05 Aug 2007 12:43:48 +0200, Charles Packer <mailbox@xxxxxxxxxxx> wrote:

Thanks for your response. The
"normalized rg" space was new to me.
I did a Google search on it and found
that the top references were concerned
with face recognition. I'm wondering
if in my images this might discard useful
information. For example, an overcast
sky could be bright in one image, dark
in another, compared to the trees.
Then the brightness would be informative.
Likewise for the blue band. I assume that
face recognition needs a fast computation
for real-time use and that's why you
can disregard that information.
For my images, I have unlimited
computation time. Anyway, you have
provided a valuable lead to more
material.


Yes, RG space is used in face detection because it is more simple to train model for segmentation using only two dimmensions. And RG is perfectly suitable for this task since skin color is mapped to very compact area in RG regardless of brightness. Of course gaussian mixture models can be used with unlimited number of dimmensions (in speech recognition they use 39 dimmension feature space!). So you can use brightness information along with color (for example HSV or YUV space). But using more dimmensions implies more complex model. That means you must use more gaussians to aproximate color distribution each class and Expectation Maximization estimation algorithm.

There is another algorithm that may work for you - Mean Shift. Look at http://www.caip.rutgers.edu/~comanici/MSPAMI/msPamiResults.html and paper http://www.caip.rutgers.edu/~comanici/Papers/MsRobustApproach.pdf. It seems more suitable for this kind of task.

Roman
.

Relevant Pages

  • Re: Best database for implementing a cache
    ... Do you need to retrieve the images ... Do you need to do face recognition, ... Sorry i think that i failed to mark the word cache in bold letter. ... What search criterion or criteria is or are used to ...
    (comp.lang.java.programmer)
  • Re: Is it error when use PCA to recognize trained face image. Why?
    ... Bo I am using the same code....for PCA and then trying to ... archive the same....to do a face recognition using ... And the train images are att face database. ...
    (comp.soft-sys.matlab)
  • Re: Best database for implementing a cache
    ... Do you need to retrieve the images ... bits and pieces of transformed image data? ... Do you need to do face recognition, ... Sorry i think that i failed to mark the word cache in bold letter. ...
    (comp.lang.java.programmer)