cy1433/Hmotion-20

---
language:
- zh
pretty_name: Hmotion-20
task_categories:
- other
tags:
- imu
- hand-motion
- motion-capture
- multimodal
- chinese
size_categories:
- 1K<n<10K
---

# Hmotion-20

Hmotion-20 is a hand-motion dataset built from wearable IMU recordings and Chinese fine-grained action annotations. The dataset is organized by subject and contains raw IMU sensor streams together with temporally aligned natural-language descriptions for left-hand and right-hand actions.

## Dataset Summary

- 20 subjects
- 400 raw IMU recording files
- 378 annotation files
- Total size: about 3.44 GiB
- Language of annotations: Chinese

The number of annotation files is smaller than the number of IMU files, which means not every IMU recording currently has a paired annotation file.

## Directory Structure

```text
Hmotion-20/
1/
IMU/
20250723091430.txt
...
label/
20250723091431_000001.txt
...
2/
IMU/
label/
...
20/
IMU/
label/
tutorial_videos/
```

## Data Format

Each subject folder contains two main subfolders:

- IMU: raw sensor recordings stored as tab-separated text files
- label: annotation files stored as JSON-formatted text files

### IMU Files

Each IMU file contains timestamped readings from wearable devices, including fields such as:

- timestamp
- device name
- acceleration on x, y, z axes
- angular velocity on x, y, z axes
- orientation angles
- magnetic field readings
- quaternion values
- temperature
- altitude and pressure when available
- firmware version and battery level

### Annotation Files

Each annotation file is a JSON object with fine-grained action descriptions for both hands. The JSON keys in the raw files are in Chinese. Typical fields include:

- left hand detailed description list
- right hand detailed description list
- timestamp
- aligned description
- refined description
- action label

## Intended Use

This dataset may be useful for:

- hand motion understanding from wearable sensors
- multimodal alignment between IMU sequences and natural-language action descriptions
- fine-grained action recognition
- sensor-based activity understanding

## Notes

- The dataset is uploaded in its original folder structure for ease of reuse.
- The tutorial_videos folder is included in the repository as part of the original release.
- Users should verify that the selected split and annotation coverage match their experimental setup.

## Citation

If you use this dataset in academic work, please cite the corresponding project or paper associated with HmotionGPT / Hmotion-20.