Datasets:

zyyy12138
/

motion-smd-data

license: other
license_name: mixed-see-subdir-readmes
language:
  - en
task_categories:
  - question-answering
  - text-generation
pretty_name: >-
  Motion-SMD: Structured Motion Descriptions and Preprocessed Motion-Language
  Data
tags:
  - human-motion
  - motion-captioning
  - motion-qa
  - biomechanics
  - llm
  - structured-text
size_categories:
  - 10K<n<100K

Motion-SMD Data

Data release for "Encoder-Free Human Motion Understanding via Structured Motion Descriptions".

🌐 Project page: https://yaozhang182.github.io/motion-smd/
💻 Code: https://github.com/yaozhang182/motion-smd
🤗 LoRA adapters: https://huggingface.co/zyyy12138/motion-smd-lora
📄 Paper (arXiv): https://arxiv.org/abs/2604.21668

What's here

Four subdirectories, each with its own README.md describing files, provenance, and license:

Subdir	Contents	Our contribution?
`smd_texts/`	Pre-generated Structured Motion Description JSONs (17 variants × 3 datasets)	Ours — derived via our rule-based pipeline
`babel_qa/`	BABEL-QA joints + original QA + our 10-option QA	Derived (joints/questions from IMoRe); 10-option QA is ours
`humman_qa/`	HuMMan-QA joints + original QA + our 10-option QA	Derived (joints/questions from IMoRe); 10-option QA is ours
`humanml3d/`	HumanML3D 22-joint sequences, texts, splits, Mean/Std	Subset mirror for reproducibility

Why this repo

The main repository (github.com/yaozhang182/motion-smd) ships only code. Reproducing the paper requires ~5 GB of preprocessed data, including SMDs computed from the three upstream datasets. This HF repo bundles all of that so git clone + huggingface-cli download is enough to reach the training step.

Quick start

pip install huggingface_hub

# Option 1: download the entire dataset (~5 GB)
huggingface-cli download zyyy12138/motion-smd-data --repo-type dataset --local-dir motion-smd-data

# Option 2: download just one subdir (e.g., SMDs only, if you already have the
# upstream data locally)
huggingface-cli download zyyy12138/motion-smd-data --repo-type dataset \
    --include "smd_texts/*" --local-dir motion-smd-data

Then symlink or move the subdirectories into the paths the code expects:

motion_smd/
├── data/babel_qa/{joints,questions,questions_10opt,angle_texts_v5*.json}
├── data/humman_qa/{joints,questions,questions_10opt,angle_texts_v5*.json}
└── captioning/data/humanml3d/{new_joints,texts,Mean.npy,Std.npy,train.txt,val.txt,test.txt,angle_texts_v5*.json}

See the root README.md of the code repo for the exact expected layout.

License

This release combines several sources under different licenses:

Our derivatives (smd_texts/*, *_qa/questions_10opt/*) are released under the MIT License.
Upstream data mirrored here (*_qa/joints/, *_qa/questions/, humanml3d/*) retains its original license. See each subdir README.

If in doubt for a given file, consult the subdir README and the upstream source linked from it.

Citation

If you use this data, please cite the paper and the upstream sources listed in each subdir.

@article{zhang2026smd,
  title   = {Encoder-Free Human Motion Understanding via Structured Motion Descriptions},
  author  = {Zhang, Yao and Liu, Zhuchenyang and Ploetz, Thomas and Xiao, Yu},
  journal = {arXiv preprint arXiv:2604.21668},
  year    = {2026}
}

Contact

Yao Zhang — yao.1.zhang@aalto.fi