README.md

EvHumanMotion

Dataset on HuggingFace 🤗

EvHumanMotion is a real-world dataset captured using the DAVIS346 event camera, focusing on human motion under diverse and challenging scenarios. It is designed to support event-driven human animation research, especially under motion blur, low-light, and overexposure conditions. This dataset was introduced in the EvAnimate paper.

Dataset Structure

The dataset is organized into two main parts:

• EvHumanMotion_aedat4: Raw event streams in .aedat4 format.
• EvHumanMotion_frame: Frame-level event slices and RGB frames.

Each is categorized into:

• indoor_day/
• indoor_night_high_noise/
• indoor_night_low_noise/
• outdoor_day/
• outdoor_night/

Each environment contains four scenarios:

• low_light/
• motion_blur/
• normal/
• over_exposure/

Example path:

EvHumanMotion_frame/indoor_day/low_light/dvSave-2025_03_04_13_02_53/
├── event_frames/
│   ├── events_0000013494.png
│   └── ...
└── frames/
    ├── frames_1741064573617350.png
    └── ...

Features

• Total sequences: 113
• Participants: 20 (10 male, 10 female)
• Duration: ~10 seconds per video
• Frame rate: 24 fps
• Scenarios: Normal, Motion Blur, Overexposure, Low Light
• Modalities: RGB + Event data (both .aedat4 and frame-level)

Applications

This dataset supports:

• Event-to-video generation
• Human animation in extreme conditions
• Motion transfer with high temporal resolution

Usage

from datasets import load_dataset

dataset = load_dataset("potentialming/EvHumanMotion")

License

Apache 2.0 License

Citation

If you use this dataset, please cite:

@article{qu2025evanimate,
  title={EvAnimate: Event-conditioned Image-to-Video Generation for Human Animation},
  author={Qu, Qiang and Li, Ming and Chen, Xiaoming and Liu, Tongliang},
  journal={arXiv preprint arXiv:2503.18552},
  year={2025}
}

Contact

Dataset maintained by Ming Li.