Research map / 3D vision / Human pose estimation
CUPS: Improving Human Pose-Shape Estimators with Conformalized Deep Uncertainty
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.
Learns a deep uncertainty score end-to-end with a GLoT-based 3D human pose-shape estimator from RGB video, then conformalizes it with weighted non-exchangeable conformal prediction and derives coverage-gap bounds.
Problem
Recovering sequences of 3D human shape and pose (SMPL) from monocular video with calibrated uncertainty and statistical coverage guarantees, where video frames violate exchangeability assumed by conformal prediction.
Approach
Generate multiple hypotheses per input during training via random frame masking, train a discriminator-style deep uncertainty function to score them, use it as the conformity score, and calibrate with Barber et al.'s weighted conformal prediction (feature-distance weights) to handle non-exchangeable data, with two miscoverage-gap bounds.
Key results
- On 3DPW (trained on 3DPW training set), CUPS reaches 48.7 PA-MPJPE, 76.2 MPJPE, 91.7 MPVPE, versus GLoT 50.6/80.7/96.3.
- Weighted conformal calibration (alpha=0.1) gives empirical coverage of 86.2% on 3DPW, 87.3% on 3DHP, and 89.0% on H3.6M, versus 81.0-85.2% for regular CP.
- Without 3DPW training, CUPS achieves 85.7 MPJPE on 3DPW versus GLoT's 89.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the deep uncertainty function and adversarial training to GLoT | GLoT · 3DPW | MPJPE: 80.7 -> 76.2 | Helped |
| add the deep uncertainty function and adversarial training to GLoT | GLoT · 3DPW | Accel: 6.6 -> 6.9 | Hurt |
| add the deep uncertainty function and adversarial training to GLoT | GLoT · MPI-INF-3DHP | MPJPE: 93.9 -> 92.8 | Helped |
| add the deep uncertainty function and adversarial training to GLoT | GLoT · Human3.6M | MPJPE: 67.0 -> 63.8 | Helped |
| add feature-distance-weighted calibration | CUPS conformal prediction · 3DPW | coverage: 81.0 ± 3.4% -> 86.2 ± 2.1% | Helped |
| use multiple randomly masked hypotheses during training (ablation) | CUPS · 3DPW | MPJPE | Helped |
| replace DUF with a classifier-style conformity score (ablation) | CUPS · 3DPW | MPJPE: 76.2 -> 77.4 | Hurt |
| add inefficiency loss to DUF (ablation) | CUPS · 3DPW | MPJPE: 76.2 -> 76.2 | No clear effect |
| set the uncertainty-loss weight to 0.6 (ablation) | CUPS training objective · 3DPW | MPJPE | Helped |
Surprising findings
- The classifier-style conformity score gives better quantitative MPJPE than no scoring function but yields less realistic predicted meshes on videos.
Limitations
- Many sampled hypotheses during training are needed, which increases GPU memory use and training time.
- The method does not use human joint-level information.
- Coverage-gap bounds rely on assumptions (e.g., periodic changepoints, Beta-distributed conformity scores) that are only approximately checked.
Experimental setup
Datasets: 3DPW Human3.6M MPII-3DHP Penn Action PoseTrack InstaVariety
Benchmarks: 3DPW Human3.6M MPI-INF-3DHP
Metrics: PA-MPJPE MPJPE MPVPE Accel empirical coverage
Baselines: VIBE MEVA TCMR MPS-Net POCO GLoT HMR 3DMB SPIN HMMR
Builds on: GLoT Barber et al. nonexchangeable conformal prediction VIBE adversarial prior learned nonconformity scores (Zhang & Carlone 2024)
Closest papers by meaning
- CHAMP: Conformalized 3D Human Multi-Hypothesis Pose Estimators (ICLR 2025 · similarity 0.88)
- Probabilistic 3D Human Shape and Pose Estimation From Multiple Unconstrained Images in the Wild (CVPR 2021 · similarity 0.88)
- PoseExaminer: Automated Testing of Out-of-Distribution Robustness in Human Pose and Shape Estimation (CVPR 2023 · similarity 0.87)
- Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.86)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.86)
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 2024 · 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.