Datasets:
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
[Paper] [Project Page] [Github]
🧙 MagicData Dataset
1. Dataset Introduction
The MagicData Dataset, introduced in MagicMotion, is a comprehensive trajectory controllable video generation dataset with rich semantic annotations. This dataset features high-quality videos with precise segmentation masks and bounding boxes annotations, designed to advance the field of controllable video synthesis and understanding.
The dataset consists of 23K diverse videos, each meticulously annotated with both pixel-level segmentation masks and object-level bounding boxes. Each video in the dataset is carefully curated to ensure high visual quality and annotation accuracy, making it suitable for training state-of-the-art video generation and understanding models.
2. File Structure
MagicData
├── videos
│ ├── videoid_1.mp4
│ ├── videoid_2.mp4
│ ├── ...
├── masks
│ ├── videoid_1
│ │ ├── annotated_frame_00000.png
│ │ ├── annotated_frame_00001.png
│ │ ├── ...
│ ├── videoid_2
│ │ ├── ...
├── boxs
│ ├── videoid_1
│ │ ├── annotated_frame_00000.png
│ │ ├── annotated_frame_00001.png
│ │ ├── ...
│ ├── videoid_2
│ │ ├── ...
├── MagicData.csv # detailed information of each video
3. Useful scripts
- Data Extraction
sudo apt-get install git-lfs
git lfs install
git clone https://huggingface.co/datasets/quanhaol/MagicData
tar -xzvf boxs.tar.gz
cat masks.tar.gz.part* > masks.tar.gz
tar -xzvf masks.tar.gz
cat videos.zip.part_* > videos.zip
unzip videos.zip
Citation
If you found this dataset useful, please cite our paper.
@article{li2025magicmotion,
title={Magicmotion: Controllable video generation with dense-to-sparse trajectory guidance},
author={Li, Quanhao and Xing, Zhen and Wang, Rui and Zhang, Hui and Dai, Qi and Wu, Zuxuan},
journal={arXiv preprint arXiv:2503.16421},
year={2025}
}
Contact
- Downloads last month
- 165