If that's the case (I think it is but correct me if I'm wrong) wouldn't it make more sense to compute the Eigen-representation of each training face first, then

If that's the case (I think it is but correct me if I'm wrong) wouldn't it make more sense to compute the Eigen-representation of each training face first, then average those Eigen face representations as opposed to the cropped faces themselves?