UMotion

Code for our paper "UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units"

Usage

Install dependencies using uv

uv venv --python 3.11
source .venv/bin/activate
uv pip install -r requirements.txt

check pyproject.toml for dependencies.

Prepare SMPL body model

1. Download SMPL model. You should download version 1.1.0 for Python 2.7 (female/male/neutral, 300 shape PCs) since UIP use neutral SMPL model.
2. Put the model file into models/smpl/. The directory structure should look like the following:

models/smpl/basicModel_f_lbs_10_207_0_v1.1.0.pkl
models/smpl/basicModel_m_lbs_10_207_0_v1.1.0.pkl
models/smpl/basicModel_neutral_lbs_10_207_0_v1.1.0.pkl

Prepare datasets

Change base_dir in src/config to your project directory.

Download the dataset and put it into datasets/raw. The directory structure should look like the following:

datasets/raw/AMASS/<subdata>/...
datasets/raw/DIP_IMU_and_Others/DIP_IMU/s_<subject_id>/...pkl
datasets/raw/TotalCapture_Real_60FPS/...pkl
datasets/raw/uip/[(test)(train)].pt

We provide different process scripts for preprocessing data into the same data structure (run in project root directory).

Dataset	Download Link	Processing Script
AMASS	https://amass.is.tue.mpg.de/ (SMPL-H)	python src/data/amass.py
DIP-IMU	https://dip-imu.github.io/ (DIP IMU AND OTHERS - DOWNLOAD SERVER 1)	python src/data/dipimu.py
TotalCapture	https://dip-imu.github.io/ (ORIGINAL TotalCapture DATA W/ CORRESPONDING REFERENCE SMPL Poses)	python src/data/totalcapture.py
UIP	https://siplab.org/projects/UltraInertialPoser	python src/data/uip.py

Run the training or evaluation

Download pre-trained models

Make sure to have gdown installed

pip install gdown

Then, run the following command line:

bash download_pretrained_models.sh

---

For shape estimator

python src/main_shape.py

For pose estimator (DIP-IMU, TotalCapture)

python src/main_pose.py # DIP-IMU, TotalCapture, without UKF (distance without noise)
python src/main_uip.py # UIP, with UKF (noisy distance)
python src/main_ukf.py # TotalCapture, with UKF (noisy distance)

Uncomment do_train() in main_shape.py and main_pose.py to train the model. The training parameters are set in configs/config.yaml and configs/config_noise.yaml

Citation

If you find this code useful in your research, please cite:

@inproceedings{liu2025umotion,
  title={UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units},
  author={Liu, Huakun and Ota, Hiroki and Wei, Xin and Hirao, Yutaro and Perusquia-Hernandez, Monica and Uchiyama, Hideaki and Kiyokawa, Kiyoshi},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={7085--7094},
  year={2025}
}

Misc

The code for sensing the distance matrix using the DW3000 is available at DW3000.

If you encounter any issues or have questions, feel free to open an issue. You may also contact me via the email address: liu.huakun.li0@is.naist.jp.