Research map / 3D vision / Human pose estimation
PostureHMR: Posture Transformation for 3D Human Mesh Recovery
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 PostureHMR, a diffusion-style 3D human mesh recovery method that progressively transforms an SMPL T-pose mesh to the target posture using a kinematics-based forward process, a mesh-to-posture decoder conditioned on a 2D mesh, and a trimap-based rendering loss.
Problem
Single-image 3D human mesh recovery (HMR) is ambiguous due to depth ambiguity, and direct image-to-posture regression or vertex-noise diffusion loses physical body structure and pixel alignment.
Approach
Replace noise-based diffusion with a posture transformation from the SMPL T-pose to the target mesh: the forward process interpolates pose via forward kinematics skeletal animation and shape linearly, and a Graphormer-based M2P decoder learns reverse steps conditioned on a 2D mesh estimated from the image, plus a trimap-weighted IoU rendering loss.
Key results
- On 3DPW, PostureHMR reports MPVE 75.4, MPJPE 64.9, PA-MPJPE 39.6, outperforming VirtualMarker (77.9/67.5/41.4) and Zolly (76.3/65.0/39.8) on the MPVE metric as reported in Table 1.
- On SURREAL, PostureHMR achieves MPVE 42.1, MPJPE 35.3, PA-MPJPE 27.4, better than VirtualMarker (44.7/36.9/28.9).
- On Human3.6M, PostureHMR gets MPVE 55.7, MPJPE 44.5, PA-MPJPE 31.0, with the best PA-MPJPE among compared methods in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace noise diffusion with vertex linear interpolation (ablation) | PostureHMR · 3DPW | MPVE (mm): 83.2 -> 80.3 | Helped |
| replace noise diffusion with vertex linear interpolation (ablation) | PostureHMR · SURREAL | MPVE (mm): 48.6 -> 46.2 | Helped |
| add forward-kinematics posture transformation (ablation) | PostureHMR · 3DPW | MPVE (mm): 80.3 -> 78.5 | Helped |
| add forward-kinematics posture transformation (ablation) | PostureHMR · SURREAL | MPVE (mm): 46.2 -> 44.5 | Helped |
| add 2D mesh conditioning (ablation) | PostureHMR · 3DPW | MPVE (mm): 78.5 -> 76.2 | Helped |
| add 2D mesh conditioning (ablation) | PostureHMR · SURREAL | MPVE (mm): 44.5 -> 42.7 | Helped |
| add trimap-based rendering loss (ablation) | PostureHMR · 3DPW | MPVE (mm): 76.2 -> 75.4 | Helped |
| add trimap-based rendering loss (ablation) | PostureHMR · SURREAL | MPVE (mm): 42.7 -> 42.1 | Helped |
Limitations
- Failure cases under occlusion or incomplete body shape produce incorrect pose estimates.
- Model-based and model-free methods are not fairly comparable due to differing backbones and training strategies, as the authors note.
Experimental setup
Datasets: Human3.6M 3DPW SURREAL MPI-INF-3DHP UP-3D COCO
Benchmarks: Human3.6M 3DPW SURREAL
Metrics: MPVE MPJPE PA-MPJPE
Baselines: HMR PC-HMR HybrIK ROMP PARE THUNDR PyMAF ProHMR OCHMR 3DCrowNet CLIFF ProPose
Builds on: VirtualMarker HMDiff MeshGraphormer Graphormer HRNet
Closest papers by meaning
- GenHMR: Generative Human Mesh Recovery (AAAI 2025 · similarity 0.88)
- ADHMR: Aligning Diffusion-based Human Mesh Recovery via Direct Preference Optimization (ICML 2025 · similarity 0.87)
- Score-Guided Diffusion for 3D Human Recovery (CVPR 2024 · similarity 0.85)
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · similarity 0.85)
- Incorporating Test-Time Optimization into Training with Dual Networks for Human Mesh Recovery (NeurIPS 2024 · similarity 0.85)
- BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation (CVPR 2025 · similarity 0.85)
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.