Research map / 3D vision / Human pose estimation
ScoreHypo: Probabilistic Human Mesh Estimation with Hypothesis Scoring
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 ScoreHypo, a diffusion-based multi-hypothesis human mesh estimator (HypoNet) paired with a pairwise-ranking hypothesis scorer (ScoreNet) that selects high-quality meshes from the generated set.
Problem
Monocular 3D human mesh estimation is ill-posed due to depth ambiguity and occlusion; probabilistic methods generate multiple hypotheses but lack a reliable way to select high-quality ones.
Approach
Formulates mesh estimation as conditional reverse diffusion over joints and twist (HypoNet) using multi-scale image features with cross-attention, then trains a separate ScoreNet with a pairwise probabilistic ranking loss on MPJPE/MPVE to score and select/average the top-K hypotheses.
Key results
- HypoNet with M=100 hypotheses reaches 38.4 MPJPE on H3.6M and 63.0 MPJPE on 3DPW, beating probabilistic baselines such as HuManiFlow (65.1 MPJPE on 3DPW at M=100).
- ScoreNet reduces MPVE on 3DPW by 13.7% for ProHMR and 17.2% for HuManiFlow relative to random selection when plugged in without fine-tuning.
- On the 3DPW-OC occlusion subset, the full method achieves 89.8 MPVE and 73.9 MPJPE, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global image feature Fg and cross-attention (ablation) | HypoNet · 3DPW test set | MPJPE: 75.9 -> 74.1 | Helped |
| add local image feature Fl (ablation) | HypoNet · 3DPW test set | MPJPE: 80.8 -> 74.1 | Helped |
| select hypotheses with ScoreNet instead of averaging them (ablation) | HypoNet · 3DPW | MPJPE: 73.9 -> 72.4 | Helped |
| select hypotheses with ScoreNet instead of random selection (ablation) | ProHMR · 3DPW | MPVPE: 120.7 -> 104.2 | Helped |
| select hypotheses with ScoreNet instead of random selection (ablation) | HuManiFlow · 3DPW | MPVPE: 112.9 -> 93.5 | Helped |
| increase generated hypothesis count from M = 10 to M = 200 (ablation) | HypoNet · H3.6M | MPJPE: 42.4 -> 37.4 | Helped |
| increase generated hypothesis count from M = 10 to M = 200 (ablation) | HypoNet · 3DPW | MPJPE: 68.5 -> 61.8 | Helped |
| use zero noise instead of random noise (ablation) | HypoNet · 3DPW | MPJPE: 78.3 -> 74.1 | Helped |
Surprising findings
- Simple averaging of HypoNet hypotheses can be no better than, or worse than, a single zero-noise sample in some settings (e.g., M=10 average 87.2 MPVE vs 86.6 zero-noise on 3DPW).
Limitations
- Relies on selecting from a fixed set of sampled hypotheses; quality of selection bounded by diversity of HypoNet outputs.
- Generalization of ScoreNet shown only by plugging into ProHMR and HuManiFlow on 3DPW; broader validation limited.
- HuManiFlow failed to produce plausible estimates in some cases, which ScoreNet cannot correct.
Experimental setup
Datasets: H3.6M 3DPW MPI-INF-3DHP MPII COCO UP-3D 3DPW-OC
Benchmarks: H3.6M 3DPW 3DPW-OC
Metrics: MPJPE PA-MPJPE MPVPE
Baselines: ProHMR HuManiFlow HMDiff Biggs et al. Sengupta et al. HybrIK CLIFF Zolly
Builds on: DDPM DDIM HybrIK ProHMR HMR
Closest papers by meaning
- Score-Guided Diffusion for 3D Human Recovery (CVPR 2024 · similarity 0.88)
- ADHMR: Aligning Diffusion-based Human Mesh Recovery via Direct Preference Optimization (ICML 2025 · similarity 0.87)
- DPMesh: Exploiting Diffusion Prior for Occluded Human Mesh Recovery (CVPR 2024 · similarity 0.86)
- FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery (NeurIPS 2026 · 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.