zeweizhang/TrajLoomDatasets

---
language:
- en
license: other
size_categories:
- 10K<n<100K
pretty_name: TrajLoomDatasets
task_categories:
- other
tags:
- trajectory
- motion
- video
license_name: trajloom-dataset-license-notice
license_link: LICENSE
---

<p align="center">
<h1 align="center"><em>TrajLoom</em>: Dense Future Trajectory Generation from Video</h1>
<div align="center">
<strong>Zewei Zhang</strong>
&nbsp;&nbsp;
<strong>Jia Jun Cheng Xian</strong>
&nbsp;&nbsp;
<strong>Kaiwen Liu</strong>
&nbsp;&nbsp;
<strong>Ming Liang</strong>
&nbsp;&nbsp;
<strong>Hang Chu</strong>
&nbsp;&nbsp;
<strong>Jun Chen</strong>
&nbsp;&nbsp;
<strong>Renjie Liao</strong>
</div>
<br>
<div align="center">
<a href="https://arxiv.org/abs/2603.22606"><img src="https://img.shields.io/badge/arXiv-Preprint-brightgreen.svg" alt="arXiv Preprint"></a>
&nbsp;
<a href="https://trajloom.github.io/"><img src="https://img.shields.io/badge/Project-Website-blue.svg" alt="Project Website"></a>
&nbsp;
<a href="https://github.com/zewei-Zhang/TrajLoom">
<img src="https://img.shields.io/badge/GitHub-Repo-black?logo=github" alt="GitHub repository">
</a>
</div>
<p align="center">
<em>More details, code, model checkpoints, and future training scripts can be found in the
<a href="https://github.com/zewei-Zhang/TrajLoom">GitHub repository</a>.</em>
</p>
</p>

# TrajLoomDatasets

TrajLoomDatasets contains the released data bundles for TrajLoom, including videos, AllTracker-processed trajectories, and cache files used for dense future trajectory generation and benchmarking.

The dataset is distributed as zip files at the repository root for easier upload and download.

## Repository layout on Hugging Face

After downloading from Hugging Face, the root folder should look like this:

```text
/TrajLoomDataset/
kubric_alltracker_s32.zip
robotap_alltracker_s32.zip
kinetics_alltracker_s32.zip
kinetics_video.zip
robotap_video.zip
kubric_video.zip
magicdata_video.zip
magicdata_alltracker_s32.zip
cache.zip
README.md
LICENSE
```

## Quick download

### Option 1: clone the full dataset repository

```bash
git lfs install
git clone https://huggingface.co/datasets/zeweizhang/TrajLoomDatasets
```

### Option 2: download with `hf`

```bash
pip install -U "huggingface_hub[cli]"
hf download zeweizhang/TrajLoomDatasets --repo-type dataset --local-dir /TrajLoomDataset
```

## Quick unzip

If the downloaded zip files are directly under `/TrajLoomDataset/`, run:

```bash
cd /TrajLoomDataset

for f in *.zip; do
unzip -oq "$f" -d /TrajLoomDataset
done
```

This restores each subset into its own root-level folder.

## Folder tree after extraction

After unzipping, the folder layout becomes:

```text
/TrajLoomDataset/
kubric_alltracker_s32/
robotap_alltracker_s32/
kinetics_alltracker_s32/
kinetics_video/
robotap_video/
kubric_video/
magicdata_video/
magicdata_alltracker_s32/
cache/
README.md
LICENSE
```

## Guided folder description

```text
/TrajLoomDataset/
├── kubric_alltracker_s32/ # AllTracker trajectory outputs for Kubric videos
├── robotap_alltracker_s32/ # AllTracker trajectory outputs for RoboTAP videos
├── kinetics_alltracker_s32/ # AllTracker trajectory outputs for Kinetics videos
├── kinetics_video/ # Kinetics video subset
├── robotap_video/ # RoboTAP video subset
├── kubric_video/ # Kubric-generated video subset
├── magicdata_video/ # MagicData video subset
├── magicdata_alltracker_s32/ # AllTracker trajectory outputs for MagicData videos
├── cache/ # Cached latents from TrajLoom-VAE from MagicData trajectory
├── README.md
└── LICENSE
```

## Notes

- The zip files are distributed at the repository root to reduce the number of uploaded files.
- The unzip command above keeps each subset in its own top-level folder.
- You may delete the `.zip` files after extraction if you no longer need them.

Optional cleanup:

```bash
cd /TrajLoomDataset
rm -f *.zip
```

## License

TrajLoomDatasets contains mixed-source materials.

- Original TrajLoom-created materials are released under Apache-2.0 unless noted otherwise.
- RoboTAP and TAP-Vid related materials follow their upstream CC BY 4.0 terms.
- Kinetics-derived materials remain subject to the original source licenses of the underlying videos.
- Kubric is Apache-2.0 software; original synthetic videos generated by us are released under Apache-2.0 unless third-party assets impose additional restrictions.

See `LICENSE` for the full dataset license notice and provenance details.

## Citation

```bibtex
@misc{zhang2026trajloomdensefuturetrajectory,
title={TrajLoom: Dense Future Trajectory Generation from Video},
author={Zewei Zhang and Jia Jun Cheng Xian and Kaiwen Liu and Ming Liang and Hang Chu and Jun Chen and Renjie Liao},
year={2026},
eprint={2603.22606},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.22606},
}
```