Systematically improving existing k-means initialization algorithms at nearly no cost, by pairwise-nearest-neighbor smoothing