Easier just to deliberately overshoot (with a too high k) and then merge any clusters with too much overlap.