Averaging multiple facial expressions through subsampling