algorithm - How does one decide the final clusters when using the means shift algorthm? -

algorithm - How does one decide the final clusters when using the means shift algorthm? -

i reading bit means shift clustering algorithm (http://en.wikipedia.org/wiki/mean_shift) , got far. each point in info set : select points within distance of (including original point), calculate mean these points, repeat until these means stabilize.

what i'm confused how 1 go here in deciding final clusters , , on conditions these means merge. also, distance used select points fluctuate through iterations or remain constant?

thanks in advance

the mean shift cluster finding simple iterative process guaranteed converge. iteration starts starting point x, , iteration steps (note x may have several components, algorithm work in higher dimensions, well):

calculate weighted mean position x' of points around x - maybe simplest form calculate average of positions of points within d distance x, gaussian function commonly used , mathematically beneficial.

set x <- x'

repeat until difference between x , x' small

this can used in cluster analysis starting different values of x. final values end @ different cluster centers. number of clusters cannot known (other <= number of points).

the upper level algorithm is:

go through selection of starting values for each value, calculate convergence value shown above if value not in list of convergence values, add together list (allow reasonable tolerance numerical imprecision)

and have list of clusters. hard thing finding reasonable selection of starting values. easy 1 or 2 dimensions, higher dimensionalities exhaustive searches not quite possible.

all starting points, end same mode (point of convergence) belong same cluster.

it may of involvement if doing on 2d image, should sufficient calculate gradient (i.e. first iteration) each pixel. fast operation mutual convolution techniques, , relatively easy grouping pixels clusters.

algorithm cluster-analysis