The algorithm was a clustering algorithm, and takes as input the number of clusters desired. The researchers put in 50 clusters, so at some point the algorithm declines to separate sparse regions, because the separations are better used elsewhere.