OpenCV - KMeans2 resulting in many empty clusters

Question

I'm using OpenCV's (Python) kmeans2 function to cluster a set of SURF features extracted from a set of images.

cv.KMeans2(mat,CLUSTER_COUNT,labels,(cv.CV_TERMCRIT_ITER,100,1.0),centers = cent)

I'm running it on a set of around 50,000 features, setting my CLUSTER_COUNT to 2000. However I only get 253 clusters which are non-empty. The remaining clusters have no features assigned to them.

I'm not sure what OpenCV's implementation of K-Means looks like, but I would have thought that when there is an empty cluster, the centre is reassigned or something?

Is there something wrong with the way I am calling it or is this the expected behaviour?

Answer 1

Maybe try cv2.kmeans instead which is easier to understand. If the problem remains please let us know.