Research map / 3D vision / Human pose estimation
GenHMR: Generative 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 GenHMR, a generative monocular human mesh recovery framework that tokenizes SMPL poses with a VQ-VAE and decodes them with an image-conditioned masked transformer, refined by 2D pose-guided latent optimization.
Problem
Monocular human mesh recovery is ill-posed due to depth ambiguity and occlusion; deterministic methods give a single output, while probabilistic multi-hypothesis methods lag behind in accuracy.
Approach
Reformulates HMR as image-conditioned masked generative modeling over discrete pose tokens, using iterative uncertainty-guided sampling (keeping high-confidence tokens and re-masking the rest) plus 2D pose-guided optimization of the decoded token embeddings at inference.
Key results
- GenHMR reports 33.5 mm MPJPE on Human3.6M, 54.7 mm on 3DPW and 68.5 mm on EMDB, versus 44.8, 70.0 and 97.8 mm for the second-best HMR2.0 in the main table.
- The claimed reductions are 25.2% on Human3.6M, 21.8% on 3DPW and 29.9% on EMDB in MPJPE relative to the best prior method.
- 2D pose-guided refinement lowers 3DPW MVE from 77.0 mm (1 iteration) to 68.5 mm (10 iterations), at an inference cost rising from 0.154 s to 0.444 s per image on an RTX A5000.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GenHMR instead of HMR2.0 | Human3.6M · Human3.6M | MPJPE: 44.8 -> 33.5 | Helped |
| use GenHMR instead of HMR2.0 | 3DPW · 3DPW | MPJPE: 70.0 -> 54.7 | Helped |
| use GenHMR instead of VQ HPS | EMDB · EMDB | MPJPE: 99.9 -> 68.5 | Helped |
| increase uncertainty-guided sampling from 1 to 20 iterations (ablation) | GenHMR · 3DPW | MPJPE: 73.5 -> 67.4 | Helped |
| increase 2D pose-guided refinement from 1 to 20 iterations (ablation) | GenHMR after UGS · 3DPW | MVE: 77.0 -> 67.5 | Helped |
| use a broader training masking ratio U(0, 0.7) rather than U(0, 0.3) (ablation) | GenHMR · 3DPW | MPJPE: 72.1 -> 67.9 | Helped |
| increase tokenizer codebook size from 1024 × 256 to 4096 × 256 (ablation) | Pose tokenizer · AMASS | MVE: 9.2 -> 5.9 | Helped |
| increase pose-token count from 48 to 384 (ablation) | Pose tokenizer · AMASS | MVE: 10.5 -> 6.0 | Helped |
| include L_2D loss (ablation) | GenHMR training · 3DPW | MPJPE: 78.5 -> 68.1 | Helped |
| remove L_3D loss (ablation) | GenHMR training · 3DPW | MVE: 77.5 -> 91.5 | Hurt |
| use a 0.8 → 0.01 temperature schedule instead of 2.0 → 0.01 (ablation) | GenHMR training · 3DPW | MPJPE: 90.1 -> 68.0 | Helped |
| extend annealing duration from 50% to 200% (ablation) | GenHMR training · 3DPW | MPJPE: 68.1 -> 75.2 | Hurt |
Surprising findings
- Narrower training masking ratios degraded accuracy, with the broadest range U(0,0.7) performing best among the tested options on 3DPW.
- Standard iterative sampling showed diminishing returns beyond about 10 iterations.
Limitations
- Inference time grows with the number of refinement iterations, trading accuracy for latency.
- Some table results are inconsistent in the text (e.g., the 20-iteration table entries and the stated error reductions), so reported numbers should be checked against the paper.
Experimental setup
Datasets: AMASS MOYO Human3.6M COCO MPI-INF-3DHP MPII InstaVariety AVA AI Challenger BEDLAM 3DPW EMDB
Benchmarks: Human3.6M 3DPW EMDB
Metrics: MPJPE PA-MPJPE MVE AITI
Baselines: FastMETRO PARE Virtual Marker CLIFF HMR2.0 VQ HPS TokenHMR Diff-HMR 3D Multibodies ProHMR
Builds on: VQ-VAE MaskGIT ViTDet Multi-scale deformable attention HMR2.0
Closest papers by meaning
- MEGA: Masked Generative Autoencoder for Human Mesh Recovery (CVPR 2025 · similarity 0.92)
- ADHMR: Aligning Diffusion-based Human Mesh Recovery via Direct Preference Optimization (ICML 2025 · similarity 0.91)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.90)
- Score-Guided Diffusion for 3D Human Recovery (CVPR 2024 · similarity 0.88)
- PostureHMR: Posture Transformation for 3D Human Mesh Recovery (CVPR 2024 · similarity 0.88)
- FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery (NeurIPS 2026 · similarity 0.88)
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.