kmeans cluster radius

edit

I think you should start by what are you trying to do, maybe there is some other approaches that fit better than kmeans...

Hi, I am trying to identify the separate clusters in images like this :

http://i58.tinypic.com/wwdw0l.jpg

So that I end up with this:

http://i59.tinypic.com/2dm9gtl.jpg

Yesterday someone suggested kmeans, which works on the still frame, but when it is a video stream, the number of clusters needs to change automatically.

So applying the kmeans on each frame is "wrong" because you have set clusterCount to 4? You'll need some sort of detector for the number of clusters...

exactly. I need a function that will find a new cluster as it appears in frame, and make it a new cluster.

Is kMeans not what i need for this? Or would i use it in conjunction with something else? What sort of detector would you recommend?

thanks for your help

Have you seen this? Have you done it in another way? Why do you want to change it? Maybe applying kmeans in a loop and detecting the best number of clusters is the way; but maybe this is not going to be fast enough.

thanks! I am unsure how to detect the best number of clusters. Will read through that wiki again and try to understand it... Possibly kMeans is not what i want though, as clusters will be constantly in and out of frame.

Is there any other function that searches for clustering in a similar way, or would you suggest another approach?

Maybe adjusting number of clusters based on the standard deviation in each cluster if you do a kmeans with clusterCount+1 and clusterCount-1?

This may help too

sorted. using the 'rule of thumb' from that wiki page, then looping and adding one each time. thanks for your help!

                    int cNumber = sqrt(centroids.size() / 2);

                    if (expectedClusterSize != actualClusterSize)
                        {
                    cNumber = (cNumber + 1);
                            kmeans(centroids, cNumber, labels, tc, attempts, flags, centers);  
                           }

...obviously will only work if all clusters are the same size.