algorithm - making clusters dynamically so that if few data points are removed other clusters remain unaffected

Question

I have a set of SIFT keypoints from a grey scale image. Is there a way that I can create clusters that do not get affected if distant keypoints are absent?

In k-means we have to decide that how many clusters we want. I am looking for method that decides K dynamically. Suppose I have n keypoints (where n is approximately 1000). Suppose the method makes 50 clusters.

Now, few keypoints are removed from one region then the method should give me same or similar clusters for areas from which keypoint was not removed.

It would be helpful if centroid of cluster undergoes very less deviation.

I tried to understand this article. Another link.

But in it, the proposed method calculates threshold and does not use it.

标签： algorithmdynamick-means