Generating Multiple 4D Expression Transitions by Learning Face Landmark Trajectories