The original method just selected the first encountered vertex in some
epsilon-neighborhood of the mouse cursor. Now the closest one is
picked. This is done in O(number_of_vertices) and could be someday
improved by more efficient nearest neighbor using k-d tree.