Double-chain Constraints for 3D Human Pose Estimation in Images and Videos