README.md

metadata

license: other
pretty_name: ProPose
tags:
  - human-pose-estimation
  - prosthesis
  - residual-limbs
  - 2d-pose-estimation
  - benchmark

ProPose Dataset

This repository hosts the dataset for ProPose: Topology-Unified 2D Pose Estimation across Intact, Residual and Prosthetic Limbs.

ProPose is a 2D human pose estimation benchmark designed for inclusive pose estimation across intact, residual, and prosthetic limbs. It follows the proposed Omni-Pose annotation protocol, which represents biological, prosthetic, and physically absent keypoints within a unified topology.

Dataset File

The dataset is released as a single ZIP archive:

ProPose.zip

After downloading and extracting the archive, the dataset is organized into training, validation, and test splits.

Dataset Structure

ProPose.zip contains the following structure:

ProPose/
├── train/
│   ├── images/
│   └── propose_train.json
├── val/
│   ├── images/
│   └── propose_val.json
└── test/
    ├── images/
    └── propose_test.json

Each split contains:

• images/: image files used for training, validation, or testing.
• propose_*.json: the corresponding pose annotation file following the Omni-Pose annotation format.

Annotation Format

The annotation files include 2D keypoint coordinates and semantic keypoint type labels. Each keypoint is associated with a semantic type:

• Bio: biological keypoint.
• Pros: prosthetic keypoint.
• Abs: physically absent keypoint.

This design enables unified pose representation for intact limbs, residual limbs, and prosthetic limbs.

Code and Documentation

The official code, training configuration, evaluation scripts, and additional documentation are available at:

https://github.com/SoraLink/ProPose

Citation

If you use this dataset or code, please cite our paper:

@inproceedings{qi2026propose,
  title={Topology-Unified 2D Pose Estimation across Intact, Residual and Prosthetic Limbs},
  author={Qi, Tianye and Zhang, Tengyue and Ying, Jiaying and Zhu, Tianqing and Yu, Xin},
  booktitle={European Conference on Computer Vision},
  year={2026}
}