Research map / 3D vision / Human pose estimation
Disentangled Diffusion-Based 3D Human Pose Estimation with Hierarchical Spatial and Temporal Denoiser
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DDHPose, a disentangled diffusion model for 2D-to-3D human pose lifting that diffuses bone lengths and bone directions separately and denoises with a hierarchy-aware spatial-temporal transformer (HSTDenoiser).
Problem
Monocular 3D human pose estimation from 2D pose sequences: disentangled bone-based methods accumulate hierarchical errors along the kinematic tree, while diffusion methods ignore explicit skeletal priors and under-model hierarchical joint relations.
Approach
Disentangle the 3D pose into bone lengths and bone directions inside the diffusion forward process (not as a separate prediction head), and add hierarchy-aware attention (HRST for parent-joint influence on spatial attention, HRTT for child-joint cross-attention in time) to the denoiser.
Key results
- On Human3.6M, DDHPose (N=243, H=20, W=10) reaches 39.0mm MPJPE, 0.5mm (1.3%) better than D3DP among probabilistic methods.
- Among disentangle-based methods, DDHPose reaches 39.7mm MPJPE vs. Anatomy3D's 44.1mm (10.0% lower).
- On MPI-INF-3DHP, DDHPose reports PCK 98.5, AUC 78.1, MPJPE 29.2mm, beating D3DP (30.2mm).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDHPose with H=20 and W=10 instead of D3DP | Human3.6M Protocol #1 · Human3.6M | MPJPE: 39.5 -> 39.0 | Helped |
| Use DDHPose instead of D3DP | MPI-INF-3DHP · MPI-INF-3DHP | PCK: 97.7 -> 98.5 | Helped |
| Use DDHPose instead of D3DP | MPI-INF-3DHP · MPI-INF-3DHP | AUC: 77.8 -> 78.1 | Helped |
| Use DDHPose instead of D3DP | MPI-INF-3DHP · MPI-INF-3DHP | MPJPE: 30.2 -> 29.2 | Helped |
| Add disentangled input (ablation) | DDHPose without disentangled input or output · Human3.6M | MPJPE: 40.23 -> 39.65 | Helped |
| Add disentangled output (ablation) | DDHPose without disentangled input or output · Human3.6M | MPJPE: 40.23 -> 41.72 | Hurt |
| Add hierarchical embedding (ablation) | DDHPose baseline · Human3.6M | MPJPE: 40.10 -> 40.08 | Helped |
| Add HRST (ablation) | DDHPose with hierarchical embedding · Human3.6M | MPJPE: 40.08 -> 39.68 | Helped |
| Add HRTT (ablation) | DDHPose with hierarchical embedding and HRST · Human3.6M | P-MPJPE: 31.55 -> 31.24 | Helped |
| Add 3D disentanglement loss (ablation) | DDHPose with 3D pose loss · Human3.6M | MPJPE: 39.87 -> 39.65 | Helped |
Surprising findings
- Disentangled output (regressing bone length and direction at the output) increased MPJPE from 40.23mm to 41.72mm, so only disentangled input was beneficial.
- Under the same setting as DiffPose's Stand-Diff, DDHPose (39.2mm) only modestly beats DiffPose's 40.1mm, and DiffPose's additional heatmap/depth priors give it a lower 36.9mm MPJPE.
Limitations
- Inference requires multiple hypotheses and iterations (H and W) to reach best numbers, increasing cost.
- Comparison with DiffPose is not like-for-like because DiffPose uses extra heatmap and depth priors.
- Hierarchy and HRST/HRTT gains are small in ablations (e.g., 0.03mm from HRTT on MPJPE).
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP
Benchmarks: Human3.6M MPI-INF-3DHP
Metrics: MPJPE P-MPJPE PCK AUC
Baselines: Anatomy3D Virtual Bones DKA VideoPose3D PoseFormer P-STMO MixSTE PoseFormerV2 STCFormer MHFormer GFPose D3DP
Builds on: D3DP MixSTE DDPM DDIM
Closest papers by meaning
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.89)
- Dual-Diffusion for Binocular 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.87)
- HiPART: Hierarchical Pose AutoRegressive Transformer for Occluded 3D Human Pose Estimation (CVPR 2025 · similarity 0.87)
- FinePOSE: Fine-Grained Prompt-Driven 3D Human Pose Estimation via Diffusion Models (CVPR 2024 · similarity 0.86)
- R-Cyclic Diffuser: Reductive and Cyclic Latent Diffusion for 3D Clothed Human Digitalization (CVPR 2024 · similarity 0.86)
- HiPoser: 3D Human Pose Estimation with Hierarchical Shared Learning at Parts-Level Using Inertial Measurement Units (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.