Research map / 3D vision / Human pose estimation
CHAMP: Conformalized 3D Human Multi-Hypothesis Pose Estimators
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 CHAMP, a diffusion-based multi-hypothesis 3D human pose estimator whose learned conformity scorer is trained end-to-end with differentiable conformal prediction to filter hypotheses before aggregation.
Problem
Single-hypothesis 3D pose estimation from 2D keypoints is ill-posed; multi-hypothesis generative estimators need a principled way to select and aggregate hypotheses, since inaccurate hypotheses skew the final output.
Approach
Learn a discriminator-style conformity score over hypotheses and optimize a differentiable conformal inefficiency loss end-to-end with a diffusion denoiser (MixSTE-based), so the model is trained to produce small, accurate conformal sets before post-hoc CP and aggregation.
Key results
- On Human3.6M with 80 hypotheses, CHAMP-Agg reaches 34.8 mm MPJPE versus 39.5 mm for D3DP and 36.9 mm for DiffPose (multi-hypothesis).
- On MPI-INF-3DHP, CHAMP-Agg achieves the best reported PCK 98.1, AUC 78.7 and MPJPE 28.6 mm among multi-hypothesis methods.
- Learned conformity score correlates with MPJPE at Pearson -0.41 (R^2 0.26 in OLS); empirical coverage on Human3.6M is around 90% ± 5% at alpha=0.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add conformal filtering before mean aggregation (ablation) | CHAMP-Naive · Human3.6M | MPJPE (Protocol #1, mm): 39.2 -> 36.9 | Helped |
| Add J-Agg aggregation while retaining conformal filtering (ablation) | CHAMP · Human3.6M | MPJPE (Protocol #1, mm): 36.9 -> 34.8 | Helped |
| Add conformal filtering before weighted mean aggregation (ablation) | CHAMP-Naive · 3DPW | PA-MPJPE (mm): 64.2 -> 61.5 | Helped |
| Replace simple aggregation with J-Agg without conformal filtering (ablation) | CHAMP-Naive · 3DPW | PA-MPJPE (mm): 64.2 -> 63.7 | Helped |
| Use the end-to-end learned conformity score instead of a separately trained or hand-designed score (ablation) | CHAMP with J-Best aggregation · Human3.6M | MPJPE | Helped |
| Increase training hypotheses above 20 (ablation) | CHAMP · Human3.6M | MPJPE | No clear effect |
| Use 80 hypotheses at inference (ablation) | CHAMP · Human3.6M | MPJPE | No clear effect |
| Set the inefficiency-loss weight to 0.6 rather than lower or higher values (ablation) | CHAMP · Human3.6M | MPJPE | Helped |
Surprising findings
- Simple mean aggregation over conformal-filtered hypotheses (CHAMP) gives about 2.3 mm gain over naive averaging of all hypotheses, while J-Agg on all hypotheses gives only 0.5 mm gain.
Limitations
- Training requires many hypotheses (20 during training), consuming large GPU memory and making training slow.
- Only single-human 3D skeleton estimation; no human shape estimation.
- Exchangeability assumption for conformal calibration is not fully satisfied on video data.
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP MPII 3DPW
Benchmarks: Human3.6M MPI-INF-3DHP 3DPW
Metrics: MPJPE P-MPJPE (Protocol #2) PCK AUC PA-MPJPE empirical coverage
Baselines: Ray3D STE P-STMO MixSTE DUE D3DP MHFormer GraphMDN DiffuPose DiffPose CameraPose 3DMB
Builds on: ConfTr D3DP MixSTE DDIM J-Agg
Closest papers by meaning
- CUPS: Improving Human Pose-Shape Estimators with Conformalized Deep Uncertainty (ICML 2025 · similarity 0.88)
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.87)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.87)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.86)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.86)
- COMPOSE: Hypergraph Cover Optimization for Multi-view 3D Human Pose Estimation (NeurIPS 2026 · 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.