Distortion-aware Motion Calibrator

Implementation of "A Self-Supervised Approach on Motion Calibration for Enhancing Physical Plausibility in Text-to-Motion".

This repository is built upon the codebase of momask-codes.

---

Requirements

The codebase has been tested with the following setup:

• Python 3.9.21
• CUDA 12.2
• PyTorch 2.1.0

---

1. Conda Environment

Create and activate a dedicated conda environment, then install dependencies:

conda env create --name dmc python=3.9
conda activate dmc
pip install -r requirements.txt

---

2. Dataset and Pretrained Models

Download HumanML3D Dataset

Download the HumanML3D dataset by following the official guidelines provided in the repository below, and place it under:

./dataset/HumanML3D

• HumanML3D repository: https://github.com/EricGuo5513/HumanML3D

We additionally provide our own mean and standard deviation statistics. These files must also be placed inside the ./dataset/HumanML3D directory.

---

Download Pretrained DMC Models

You can download the pretrained Distortion-aware Motion Calibrator (DMC) models using:

bash prepare/download_models.sh

If the script fails, the models can be downloaded manually from the following link:

• https://drive.google.com/file/d/1euhjd7IPoLb_TbCg4rwic9U8gp62OcYF/view

---

3. Additional Dependencies

Evaluation Models

Download the required evaluation models and GloVe embeddings:

bash prepare/download_evaluator.sh
bash prepare/download_glove.sh

---

Training

Train WGAN-based DMC

python train_dmc_wgan.py --name wgan --config_path ./config/wgan.yaml

Train Denoising-based DMC

python train_dmc_denoising.py --name denoising --config_path ./config/denoising.yaml

---

Evaluation

Evaluate WGAN-based DMC

python eval_dmc_wgan.py --name dmc_wgan --config_path ./config/wgan.yaml --save_anim

Evaluate Denoising-based DMC

python eval_dmc_denoising.py --name dmc_denoising --config_path ./config/denoising.yaml --save_anim

---

Visualization

Motion results are rendered using Isaac Sim.

To visualize 3D joint sequences in Isaac Sim, we use the following renderer:

• https://github.com/ahrilab/SMPL-renderer

This renderer converts motion outputs into Isaac Sim-compatible visualizations.

---

Acknowledgements

This project builds upon the implementation and dataset ecosystem provided by:

• https://github.com/EricGuo5513/momask-codes
• https://github.com/EricGuo5513/HumanML3D