rww@esl.UUCP (Richard W. Webb) (04/21/88)
Hello,
I am working on a personal project that needs to be able to
automatically find clusters in a set of data points in N dimensions.
This will be used to iteratively refine the interesting areas in
certain Fractal objects. I have also considered using it in a
hierarchical rendering system. Given a set of M points, I would
like any information relevant to the problem of associating points
with one local adjacent.
One method I have looked at is to make a Minimum spanning tree
for the points and throwing away the longest branches until I have
nice clusters. This method does work nicely, but has to have some
arbitrary hueristics to decide when to stop.
I have heard of a program called ISODATA that does some clustering,
but I don't know how to get it or what really does.
Any help or SOURCE :-) would be greatly appreciated.
--
Richard W. Webb ecvax!decwrl!borealis!\
ESL Incorporated sdcsvax!seismo!- ames!esl!rww
ARPA: rww%esl@ames.ARPA ucbcad!ucbvax!/ /
SMAIL: rww@esl.ESL.COM ihnp4!lll-lcc!finegan@ucqais.uc.edu (Mike Finegan) (04/25/88)
In article <655@esl.UUCP>, rww@esl.UUCP (Richard W. Webb) writes: > Hello, > > I have heard of a program called ISODATA that does some clustering, > but I don't know how to get it or what really does. > I used ISODATA to cluster 3-D vectors (RGB, HSI, etc.) for image processing segmentation. The authors are Ball & Hall - Stanford Research Labs (sic?). The algorithm is presented fairly completely in the report (dated 1965). The parameters which control the clustering are cryptic (i.e. heuristic ...) and it uses large amounts of memory. It worked okay for my application, but I didn't have the patience to try every combination of process parameters, to see if I was getting optimum clustering or not. You would probably be better off with another technique, unless you have very few data points (or not too many dimensions). If you really want code - I could give you what I have - it was written for a class and could be cleaned up. Maybe you could send me some other clustering algorithms in return ? I have a copy of the report - its about 100 pages - but the algorithm is about 10 or 20, with flow charts, diagrams, formulae, etc. Mike Finegan ...uccba!finegan