(updated June 30)
I used distruct to create a graphical display of the clusters revealed by my 2004 model-based clustering of 2,504 human skulls on 57 metrical variates. As mentioned in the original article, these traits are enough to distinguish some human races (e.g., Caucasoids=Norse, Zalavar, Berg, Egyptian), or even individual populations (e.g., Eskimo, Buriat, or Bushmen).
Of course, some skulls don't fall in the right cluster, but this is to be expected both due to the state of the original collections (*) and due to the plasticity of the human skull that may create false associations.
But, on the whole, the clusters emerge as distinct and unmistakable entities; this level of resolution at a global scale is only possible -if at all- with hundreds of thousands of genetic markers, yet 57 linear measurements pretty much do the same trick.
One can only imagine what would be possible if someone takes a 3D scanner around the world to visit the same museum collections that Howells did several decades ago. But, perhaps, physical anthropologists have better things to do these days than discovering differences between human populations...
(*) For example, W.W. Howells noted in his work that one of the American skulls obviously belonged to a white settler.
UPDATE (June 30)
Please consult the original article for details on the populations and the methodology used. Note that K=14 is the number of clusters which maximizes the Bayes Information Criterion, but it is by no means the end of the story. For even higher K, some populations can be further separated, although some of them (e.g., Europeans) never split into fairly "clean" clusters with these 57 variables.
Below are all the results for K=2 to K=14. As with Rosenberg et al. (2002) and the work that followed it, the first split contrasts East Asians with Eurafricans. It is important to note that the pattern of successive splits should not be interpreted as a phylogeny of human populations, i.e., a history of human subdivisions.