Improving Robustness and Accuracy via Relative Information Encoding in 3D Human Pose Estimation