Bridging 2D Observations and Learned 3D Kinematics: Target-Space Anchoring and Refinement for 3D Human Pose Estimation

Mamba holds great promise for modeling long-range spatiotemporal dependencies in monocular 3D human pose estimation (3D-HPE). However, existing Mamba-based methods generally propagate states over representations derived from 2D poses, where 3D skeletal and trajectory patterns are not explicitly represented during state modeling. To address this limitation, we propose KAR-Mamba, a kinematic anchoring and refinement framework that learns target-space kinematic representations and explicitly uses them to guide Mamba state propagation and updating. First, Codebook-Guided 3D Pose Anchoring leverages learned skeletal and trajectory codebooks to correcte initial 3D relations toward representative human motion patterns and construct a kinematically informed pose anchor. Then, Skeletal-Trajectory Aligned Scanning organizes state propagation along skeletal chains and joint trajectories, with anchor-derived kinematic dependencies aligned to the corresponding transitions. Kinematic-Dependency-Modulated State Updating uses these relations to modulate input-dependent state parameters. Finally, the refined representations are decoded into anchor-relative offsets to predict the 3D pose. Experiments on Human3.6M and MPI-INF-3DHP show competitive accuracy and support the usefulness of incorporating learned 3D kinematic representations into refinement. Our code will be made publicly available.