SCALE-Pose: Skeletal Correction and Language Knowledge-assisted for 3D Human Pose Estimation

Transformer-based 3D human pose estimation methods typically use 2D joint sequences as inputs, leveraging spatial and temporal transformer encoders to model the 3D human pose. However, these methods often neglect to incorporate skeletal constraints to limit joint motion, and few consider integrating prior category knowledge to enhance potential joint representations. To solve these problems, we propose a new method named SCALE-Pose. Firstly, this method incorporates the spatial and temporal skeleton correction blocks to improve the ability of modeling the long-range dependency of the spatiotemporal motion of specific skeletons. Next, a four-stream radian loss based on skeleton angle error is introduced to constrain the motion space of joints. Finally, an auxiliary method employs global-local prompts from a large language model to generate prior category knowledge, improving the ability to generalize prior category knowledge. Experimental results on Human3.6M and MPI-INF-3DHP datasets demonstrate that our method outperforms existing approaches.