Answers are grounded in this paper and cite their sources.
Video-based 3D human pose estimation suffers from depth ambiguity and faces great challenges in modeling long‑range spatial and temporal dependencies. Diffusion models have emerged as a powerful paradigm for this task. However, their denoisers commonly rely on Transformer backbones, where quadratic self-attention complexity limits computational efficiency. We propose AnatoMamba, an anatomy-aware Mamba framework for diffusion-driven 3D human pose estimation. AnatoMamba employs Articulated Pyramid Pooling to aggregate multiscale body context, Kinematic Wave Scanning to serialize joints according to skeletal topology, and Anatomical Signature Injection to explicitly encode body-part membership. The derived anatomy-aware features are further refined through bidirectional spatial and temporal state space modeling to capture inter-joint dependencies and cross frame motion patterns. With 243-frame inputs and 10.44 million trainable parameters, AnatoMamba-L achieves an MPJPE of 38.5 mm on Human3.6M. In cross-dataset evaluation on MPI-INF-3DHP, it achieves 98.9% PCK, 87.5 AUC, and 14.7 mm MPJPE. These results demonstrate competitive pose estimation accuracy with a compact parameter footprint.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: