I''ve been trying to figure out what the following two terms actually mean. I''ve found many references and test results for each, but no actual description of the techniques... Without further delay: What is the mean-shift algorithm? How does it work? What does it do? It''s not just a regular mean-filter right? What is Kernel Density Estimation (specifically the Parzen Window technique). How do these things work? Like I said, I''ve found many references to these techniques, but I need a good desription. Im very interested in learning more about this area of research, so any links to similar topics would be greatly apreciated. Thanks very much, Will

Why are you interested in these things? Keep in mind that this forum is focused on game development issues!

Graham Rhodes
Senior Scientist
Applied Research Associates, Inc.

Hi,
[Following methods are used for classification of data]

Parzen window and the closely related K-nearest neighbour (K-NN) are a part of methods known as non-parametric-techniques.
That means that they are used when no model is assumed. That is, a histogram is a simple nonpara. They are used to estimate densities of data given some datasamples.

Both parzen and knn approximate the density by the ratio of datasamples within a volume to total datasamples (withinVol/total). The difference is that for parzen the VOLUME is fixed and k (sampleswithinvol) is determined from data. Whereas knn use a fixed K and determines the corresponding volume (increasing/decreasing to enclose k).

Anyway, look up Density estimation (both parametric and nonparam.). I'm sure there are loads of books on this, and on the net!

PS. parzen is part of the techniques called Kernel Density estimation, KDE.

/regards

/Mankind gave birth to God.

[edited by - silvren on April 18, 2003 4:09:12 PM]