Research map / 3D vision / Human pose estimation
On the Calibration of Human Pose Estimation
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.
Shows theoretically and empirically that heuristic confidence in 2D human pose estimation is miscalibrated w.r.t. OKS, and proposes CCNet, a light post-hoc branch trained on penultimate features to predict OKS-aligned keypoint confidence and visibility, improving COCO AP across pose models.
Problem
Confidence scores in 2D human pose estimation are heuristic (heatmap max or RLE uncertainty) and do not align with the OKS accuracy measure that mAP depends on, so the ranking used for AP is miscalibrated.
Approach
Derives the ideal expected-OKS confidence under a Gaussian annotation-noise assumption, showing heatmap and RLE confidences have scaling and form gaps; then attaches a lightweight ConfidenceNet on frozen penultimate features, supervised by per-keypoint OKS (MSE) and ground-truth visibility (BCE), to learn network-specific calibrated confidence.
Key results
- On COCO val, CCNet improves mAP over off-the-shelf models by up to 1.4 points (RLE ResNet-50: 72.2 to 73.6), e.g. HRNet-W32 76.0 to 77.0 and ViTPose-B 77.3 to 78.1.
- On 3DPW, adding CCNet to SPIN+SBL lowers MPJPE from 100.5 to 99.7 and PA-MPJPE from 58.8 to 57.8.
- On MPII, CCNet raises PCK-based mAP for RLE from 75.4 to 76.6 and for SBL from 77.3 to 77.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CCNet | SBL with ResNet-50 · COCO val | mAP: 72.4 -> 73.3 | Helped |
| Add CCNet | RLE with ResNet-50 · COCO val | mAP: 72.2 -> 73.6 | Helped |
| Add CCNet | HRNet-W32 · COCO val | mAP: 76.0 -> 77.0 | Helped |
| Add CCNet | Poseur with ResNet-50 · COCO val | mAP: 76.8 -> 77.7 | Helped |
| Add CCNet to part confidence | Poseur · COCO-WholeBody val | whole-body mAP: 61.0 -> 63.3 | Helped |
| Add CCNet | RLE · MPII val | mAP: 75.4 -> 76.6 | Helped |
| Replace estimated confidence with a constant (ablation) | SBL · COCO val | mAP: 72.4 -> 67.5 | Hurt |
| Adjust confidence using predicted area (ablation) | SBL · COCO val | mAP: 72.4 -> 73.0 | Helped |
| Set confidence to OKS (ablation) | SBL · COCO val | mAP: 72.4 -> 75.6 | Helped |
| Replace estimated confidence with a constant (ablation) | RLE · COCO val | mAP: 72.2 -> 67.2 | Hurt |
| Adjust confidence using predicted area (ablation) | RLE · COCO val | mAP: 72.2 -> 73.2 | Helped |
| Set confidence to OKS (ablation) | RLE · COCO val | mAP: 72.2 -> 75.4 | Helped |
Surprising findings
- Replacing predicted confidence with a constant value drops mAP by about 5 points, showing confidence strongly affects AP.
- Sophisticated uncertainty-estimation losses (aleatoric, deep ensembles, SOAP) perform no better than plain MSE for post-hoc confidence calibration.
Limitations
- Analysis assumes a perfectly trained network with Gaussian ground-truth annotation noise, so closed-form correction alone cannot fully close the gap.
- Calibration is post-hoc and depends on the frozen pose network's features; pre-training confidence estimation is observed to hurt accuracy.
- Gains on MPII and downstream 3D fitting are evaluated only on a limited set of models.
Experimental setup
Datasets: MSCOCO MPII MSCOCO-WholeBody 3DPW
Benchmarks: MSCOCO MPII MSCOCO-WholeBody 3DPW
Metrics: mAP mAR AP50 AP75 PCK AUSE Pearson correlation MPJPE PA-MPJPE MVE
Baselines: SBL HRNet ViTPose RLE IPR Poseur SPIN
Builds on: RLE SBL Poseur CLIFF
Closest papers by meaning
- 3D Human Pose Estimation = 2D Pose Estimation + Matching (CVPR 2017 · similarity 0.87)
- 3D Human Pose Estimation From a Single Image via Distance Matrix Regression (CVPR 2017 · similarity 0.87)
- LCR-Net: Localization-Classification-Regression for Human Pose (CVPR 2017 · similarity 0.87)
- ProbPose: A Probabilistic Approach to 2D Human Pose Estimation (CVPR 2025 · similarity 0.85)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.85)
- PoseIRM: Enhance 3D Human Pose Estimation on Unseen Camera Settings via Invariant Risk Minimization (CVPR 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.