README.md

---
pretty_name: T2exture Dataset
task_categories:
- image-to-image
tags:
- thermal-video
- lwir
- frame-interpolation
- texture-propagation
gated: true
extra_gated_heading: "Request access"
extra_gated_description: "Please request access before downloading this dataset. If you use the dataset, you must cite the T2exture paper."
extra_gated_button_content: "Request access"
---

# T2exture Dataset

Prepared synthetic and real thermal sequences for the two-stage T2exture method.

## Access

This dataset uses Hugging Face gated access. Please request access before downloading. No additional institution or intended-use form is required.

## Citation requirement

If you use this dataset in a paper, thesis, report, benchmark, software release, model card, or public result, you must cite:

> Jiashuo Chen, Cheng Dai, Yanan Hu, and Fanglin Bao. *T2exture: Sparsely Perturbed Thermal-to-Texture Imaging*. arXiv:2608.02192, 2026. DOI: 10.48550/arXiv.2608.02192.

BibTeX:

```bibtex
@article{chen2026t2exture,
  title   = {T$^2$exture: Sparsely Perturbed Thermal-to-Texture Imaging},
  author  = {Chen, Jiashuo and Dai, Cheng and Hu, Yanan and Bao, Fanglin},
  journal = {arXiv preprint arXiv:2608.02192},
  year    = {2026},
  doi     = {10.48550/arXiv.2608.02192},
  url     = {https://arxiv.org/abs/2608.02192}
}
```

Please also report the dataset URL and release commit used in your experiment:

`https://huggingface.co/datasets/chenjiashuo/T2exture_datasets`

The machine-readable citation is provided in [`CITATION.cff`](CITATION.cff).

## Download

After access is approved:

```bash
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='chenjiashuo/T2exture_datasets', repo_type='dataset', local_dir='datasets')"
```

## Layout

```text
datasets/
  train.txt
  valid.txt
  test.txt
  dataset_manifest.json
  sim/<scene>/
    texture/001.npy       # prepared target residual X
    passive/001.npy       # passive/source-off state S^off
    source_on/001.npy     # optional raw S^on for exact anchor construction
  flow/s10/<scene>/001_002_011.npz
  source_off/amt-s/<scene>/001.npy
  source_off/amt-l/<scene>/001.npy
  source_off/amt-g/<scene>/001.npy
  real/<sequence>/...
```

Synthetic frames are single-channel NumPy arrays. Real benchmark frames are grayscale PNG images. The prepared synthetic files store `X` directly in `texture/`; `source_on/` is optional and is copied when present by the dataset preparation script.

## Protocol

The default protocol uses an active-frame stride of `10`, a centered passive context of `5`, and `flow/s10`:

```text
C_t = [S^off_{t-2}, S^off_{t-1}, S^off_t, S^off_{t+1}, S^off_{t+2}]
```

Formal T2exture-S/L/G runs use the matching Stage 1 cache:

```text
T2exture-S -> source_off/amt-s
T2exture-L -> source_off/amt-l
T2exture-G -> source_off/amt-g
```

These caches contain estimated `S_hat^off` frames at active keyframes. Stage 2 uses them both to correct active texture anchors and to fill the passive context at active keyframes.

## Local Check

After downloading this dataset as `datasets/`, run from the code repository root:

```bash
python -B scripts/preflight.py --data-root datasets --config train.yaml --require-source-off
```

The check verifies split files, frame shape, flow files, and all three Stage 1 cache folders.

## Excluded Files

Training logs, evaluation outputs, generated visual sheets, previews, and local caches are not dataset artifacts and should stay outside this repository.