MvP: Learning Motion Prior for Diffusion-Based 3D Human Pose Estimation

Recent diffusion-based solutions for 3D Human Pose Estimation (HPE) leverage generative modeling to address depth ambiguity and uncertainty in 2D-to-3D lifting. While substantial progress has been made in spatial accuracy, the generation of temporally coherent trajectories remains largely underexplored, which is crucial for understanding realistic human motion. In this work, we propose to model a motion prior for temporally consistent joint trajectories, together with a pose prior for anatomically plausible 3D poses, through independent diffusion models. Specifically, we represent motion using relative joint displacements, defined as the positional differences of joints between frames. The motion prior thus refers to the distribution of relative joint-displacement sequences modeled by the displacement denoiser. Motivated by the observation that human joint motion is typically smooth over short time periods, we introduce a keyframe denoiser to estimate poses at uniformly sampled keyframes, which also serve as anchors for displacement accumulation. Accordingly, we decompose 3D HPE into keyframe estimation and displacement estimation. Two specialized denois ers with exchanged attention orders and independent objec tives are designed to capture spatial correlations and temporal dependencies, respectively. We further introduce an endpoint corrected pose aggregation scheme that accumulates predicted displacements while maintaining consistency between adjacent keyframes. Extensive evaluations on three widely used 3D HPE benchmarks, Human3.6M (H36M), MPI-INF-3DHP (3DHP), and the challenging in-the-wild 3DPW, demonstrate the effectiveness of our method in balancing spatial accuracy and temporal smoothness. Notably, MvP achieves competitive spatial accuracy under zero-shot evaluation and improved temporal consistency after adaptation to the target domain.