vrg-prague/ProbPose-s

📦 ProbPose: Probabilistic Human Pose Estimation

ProbPose introduces a probabilistic framework for human pose estimation, focusing on reducing false positives by predicting keypoint presence probabilities and handling out-of-image keypoints. It also introduces the new Ex-OKS metric to evaluate models on false positive predictions.

📝 Model Details

• Model type: ViT-s backbone with ProbPose head
• Input: RGB images (192x256)
• Output: Coordinates, uncertainties, quality and visibility for human keypoints
• Language(s): Not language-dependent (vision model)
• License: GPL-3.0
• Framework: MMPose

🧠 Training

• Training data: COCO Dataset
• Training script: GitHub - ProbPose_code
• Epochs: 210
• Batch size: 64
• Learning rate: 5e-5
• Hardware: 4x NVIDIA A-100

📈 Evaluation

• Metrics: mAP and Ex-mAP
• With GT bounding boxes

Dataset	mAP	Ex-mAP
COCO	76.6	76.4
CropCOCO	81.7	73.9
OCHuman	60.4	60.2

📄 Citation

If you use ProbPose in your research, please cite:

@inproceedings{probpose2025,
  title={{ProbPose: A Probabilistic Approach to 2D Human Pose Estimation}}, 
  author={Miroslav Purkrabek and Jiri Matas},
  year={2025},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
}

🧑‍💻 Authors

• Miroslav Purkrabek (personal website)
• Jiri Matas (personal website)