Research map / 3D vision / Human pose estimation
ProbPose: A Probabilistic Approach to 2D 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.
Proposes ProbPose, a top-down 2D pose estimator that outputs calibrated probability maps, a presence probability for out-of-window keypoints, and visibility, plus a CropCOCO dataset and Ex-OKS metric for out-of-image keypoints.
Problem
Top-down human pose estimators ignore out-of-image keypoints in training and evaluation and rely on uncalibrated Gaussian heatmaps, so they cannot indicate whether a keypoint is present in the activation window or express calibrated uncertainty.
Approach
Replace heatmaps with sparsemax-normalized probability maps trained with a pixel-wise OKS expected-risk loss, decode by maximizing expected OKS, add a separately trained presence probability (BCE) for keypoints outside the activation window, and generate out-of-image training samples via crop augmentation.
Key results
- On CropCOCO, ProbPose-s reaches 81.7 mAP and 73.9 Ex-mAP versus 72.7 mAP and 66.5 Ex-mAP for ViTPose-s.
- On COCO, ProbPose-s achieves 76.6 mAP versus 75.9 for ViTPose-s, with Ex-mAP 76.4 versus 75.3.
- Presence probability reduces classification error relative to ViTPose confidence by 30% (unbalanced) and 45% (balanced) on CropCOCO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add crop data augmentation (ablation) | ProbPose-s · CropCOCO | mAP: 72.4 -> 81.7 | Helped |
| add crop data augmentation (ablation) | ProbPose-s · COCO | mAP: 76.0 -> 76.6 | Helped |
| add probability maps without crop augmentation (ablation) | ProbPose-s · COCO | mAP: 76.0 -> 76.4 | Helped |
| add probability maps without crop augmentation (ablation) | ProbPose-s · CropCOCO | mAP: 73.7 -> 72.4 | Hurt |
| add probability maps alongside crop augmentation (ablation) | ProbPose-s · COCO | mAP: 76.6 -> 76.6 | No clear effect |
| add probability maps alongside crop augmentation (ablation) | ProbPose-s · CropCOCO | mAP: 81.7 -> 81.7 | No clear effect |
| use the double-probmap approach (ablation) | ProbPose-s · COCO | mAP: 76.6 -> 76.2 | Hurt |
| use the double-probmap approach (ablation) | ProbPose-s · CropCOCO | mAP: 81.7 -> 80.9 | Hurt |
| use the double-probmap approach (ablation) | ProbPose-s · OCHuman | mAP: 60.4 -> 61.4 | Helped |
| replace ViTPose-s with ProbPose-s | ViTPose-s · COCO | mAP: 75.9 -> 76.6 | Helped |
| replace ViTPose-s with ProbPose-s | ViTPose-s · CropCOCO | mAP: 72.7 -> 81.7 | Helped |
| replace ViTPose-s with ProbPose-s | ViTPose-s · OCHuman | mAP: 60.3 -> 60.4 | Helped |
Surprising findings
- Probability maps alone (without crop augmentation) reduce out-of-image accuracy on CropCOCO compared with heatmaps, while giving negligible COCO gains.
- COCO keypoints near bounding-box borders are often annotated further inside the image, penalizing accurate out-of-image localization; excluding them shows about 0.2% difference.
Limitations
- Probability maps do not outperform heatmaps on in-image keypoints; gains are about 1%, comparable to other augmentations.
- Behavior of probability maps in multibody scenarios is left for future work.
- Annotation noise near bounding-box borders in COCO limits measured out-of-image accuracy and is not resolved.
Experimental setup
Datasets: COCO OCHuman
Benchmarks: COCO CropCOCO OCHuman
Metrics: mAP Ex-mAP Ex-OKS classification accuracy for presence probability
Baselines: ViTPose-s HRFormer-s PVTv2 SWIN-t RTMO Hide-and-Seek
Builds on: ViTPose OKSLoss UDP decoding Hide-and-Seek Sparsemax
Closest papers by meaning
- Distribution-Aware Coordinate Representation for Human Pose Estimation (CVPR 2020 · similarity 0.89)
- SHaRPose: Sparse High-Resolution Representation for Human Pose Estimation (AAAI 2024 · similarity 0.87)
- MV-SSM: Multi-View State Space Modeling for 3D Human Pose Estimation (CVPR 2025 · similarity 0.87)
- Optimizing Human Pose Estimation Through Focused Human and Joint Regions (AAAI 2025 · similarity 0.87)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.87)
- Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (NeurIPS 2024 · 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.