🚀 Welcome to RUMPL

June 1, 2026 · View on GitHub

RUMPL: Ray-Based Transformers for Universal Multi-View 2D to 3D Human Pose Lifting

Hi there! 👋 Welcome to the official repository for RUMPL. We are excited to share that our paper has been accepted to IEEE Transactions on Image Processing (TIP)!

This framework is designed to help you with 3D pose estimation and motion learning using the power of your dataset. Below is everything you need to get set up and running smoothly. Let's dive in! 🏊‍♂️

📢 News

[2026] RUMPL has been accepted to IEEE Transactions on Image Processing (TIP). You can find the updated citation details below.

🛠️ Installation

Getting started is easy. Just follow these steps to prepare your environment.

1. Install Python Dependencies

First things first, let's grab the necessary python packages:

pip install -r requirements.txt

2. External Tools

RUMPL relies on two awesome external frameworks. You'll need to install them separately:

🏃 AMASS Framework: Required for running MHP. 👉 Install AMASS here
📸 MMPose: Required for off-the-shelf 2D pose estimation. 👉 Install MMPose here

📦 Dataset Setup

You will need the AMASS dataset for training (plus camera calibrations from your test dataset). RUMPL has been tested on CMU, Human3.6M, and RICH.

Here is how to prep each one:

🏛️ CMU Panoptic

Download the images and annotations using the Panoptic Toolbox.

Once downloaded, correct the paths in our script and run:

cd RUMPL/data
sh preprocess_cmu_panoptic_all_cams.sh

🧍 Human3.6M (H3.6M)

Prepare images and annotations using the H36M-Toolbox.
Update your paths and run:
```
cd RUMPL/data
sh preprocess_h36m.sh
```

💰 RICH Dataset

Download and prepare data from the RICH Website.
Update your paths and run:
```
cd RUMPL/data
sh preprocess_rich.sh
```

🏃 Generating 3D Data (MHP)

To generate the 3D dataset for RUMPL, we use the MHP module.

Download SMPL Data: Head over to the AMASS Website and download the pose SMPL data (SMPL+H is supported).
Run the Generation: We have prepared scripts for each dataset. Make sure you have the amass framework installed and your paths configured.

    cd MHP
    sh run_mmpose_00_cmu.sh      # For CMU
    sh run_mmpose_00_h36m.sh     # For Human3.6M
    sh run_mmpose_00_rich.sh     # For RICH
    ```

    > **💡 Pro Tip:** Take a peek inside these scripts before running them! There are some parallel processing commands that are commented out—you might want to enable them to speed things up.

🙌 Acknowledgements

We stand on the shoulders of giants! 🌍 Big thanks to the authors of these amazing projects:

PPT
PoseFormer
H36M-Toolbox (Microsoft)

⚖️ License & Copyright

This project is licensed under the GNU Affero General Public License v3.0 (AGPLv3).

This means you are free to use, modify, and distribute this software, but if you run it as a network service (SaaS), you must share your source code.

See the LICENSE file for the full legal text.

Happy Coding! 💻✨ If you have questions, run into bugs, or just want to say hi, feel free to open an issue. We're here to help!

📝 Citation

If you find this code useful for your research, please consider citing our paper:

@ARTICLE{11534395,
  author={Ghasemzadeh, Seyed Abolfazl and Alahi, Alexandre and De Vleeschouwer, Christophe},
  journal={IEEE Transactions on Image Processing}, 
  title={RUMPL: Ray-Based Transformers for Universal Multi-View 2D to 3D Human Pose Lifting}, 
  year={2026},
  volume={35},
  number={},
  pages={5509-5522},
  keywords={Cameras;Training;Modeling;Transformers;Testing;Computer vision;Computers;Conferences;Pose estimation;Learning (artificial intelligence);Multi-view;3D human pose estimation;ray representation;transformer;generalization},
  doi={10.1109/TIP.2026.3694127}}