To gain efficiency, we should switch from (R,t) representation for Transform3 to leverage CPU vectorization.