Geometry clustering is usually useful when an object is supposed to be represented by bill-boards, which is a common technique used in real-time graphics.
There are a bunch of clustering algorithms, and "k-means" is one of them, and a relatively simple one, in my opinion. According to "http://en.wikipedia.org/wiki/K-means_clustering", clustering is a NP-hard problem even in 2D plane, so only approximation methods are available. k-means, adopts heuristic strategy to solve it. It seeds by selecting several random triangles(or quad, etc.), then go through each of the rest ones to see which current cluster is the "closest" one [this weighting is what I am investigating for something new], then affiliate it to this cluster, then update the merged cluster.
The original version of k-means just uses space distance as the weighting function, like this:
As the above model "Spruce" shows, more than 12k triangles have been clustered into 8 groups.