Part-Aware Structured Register Modeling for 3D Human Pose Estimation

Current 2D-to-3D lifting approaches have achieved notable success in 3D human pose estimation. However, relying solely on 2D joint sequences leaves these methods susceptible to 2D detection noise and overlooks rich visual cues in RGB images that can help resolve depth ambiguities. Recent visually assisted pose lifting methods attempt to leverage such visual information, but largely focus on localized regions around predicted joints, overlooking high-level semantic visual representations and hierarchical structural context of the human body. To address these limitations, we propose Part-Aware Structured Register Modeling for 3D Human Pose Estimation (PSRM), a new framework comprising Global-to-Part Expert Adapter (GPEA) and Kinematic Register Attention (KRA). To bridge the granularity mismatch between frame-level visual context and joint-level pose features, GPEA progressively adapts and integrates visual cues through a cascade of global and part-specific experts. In addition, KRA introduces persistent global and part-level registers that iteratively interact with joint representations under kinematic constraints, enabling the aggregation and layer-spanning preservation of structural context. Extensive experiments on Human3.6M and MPI-INF-3DHP demonstrate that PSRM achieves state-of-the-art performance. Moreover, GPEA and KRA feature a highly versatile plug-and-play design, seamlessly integrating into representative pose lifting backbones and consistently boosting their precision.