Efficiently prune and recover pose tokens to improve the efficiency of transformer-based 3D human pose estimation.