---
pretty_name: Comprehensive Flame Dataset - Validation Split
task_categories:
- object-detection
size_categories:
- 1K<n<10K
tags:
- fire
- flame
- fire-detection
- yolo
- pascal-voc
- validation
---

# Comprehensive Flame Dataset - Validation Split

This repository provides the existing validation split of the **Comprehensive Flame Dataset** (`fire_rgb`), the visible-light flame dataset associated with the [Small-Scale Fire Datasets project](https://github.com/yang1002-life/small-fire_datasets) and the paper [FireDet: A framework for precise detection and localization of small-scale fire](https://doi.org/10.1016/j.firesaf.2026.104737).

Only the original validation split is provided here. No training split is included, and no new split has been generated. The counts below describe this release, rather than the totals for other datasets described by the paper or project.

## Contents

| Content | Count |
| --- | ---: |
| Validation images | 2,424 |
| JPEG images | 2,133 |
| PNG images | 291 |
| YOLO TXT label files | 2,424 |
| Original XML annotation files | 2,424 |
| Fire bounding boxes in YOLO labels | 3,231 |
| Fire bounding boxes in XML annotations | 3,231 |
| Additional smoke bounding boxes in XML annotations | 576 |
| Empty YOLO label files | 899 |

```text
images/val/         # Original validation images
labels/val/         # Original YOLO labels; class 0 = fire
annotations/val/    # Original XML annotations; fire and smoke
manifest.json      # Relative paths, byte sizes and SHA-256 of all 7,272 data files
SHA256SUMS         # SHA-256 checksums of the same data files
README.md
```

Images and annotation files are preserved byte for byte, including their original filenames. Operating-system thumbnails and training caches are excluded. The raw images and annotations together occupy approximately 2.14 GB (decimal).

## Annotation formats

Each image is paired with one TXT and one XML file with the same filename stem.

**YOLO TXT:** each nonempty line is `0 x_center y_center width height`, with coordinates normalized to image dimensions. The single class is `0: fire`. Empty TXT files are intentional: these images have no fire boxes in the supplied labels and are retained for validation. An empty fire label does not imply that an image has no smoke.

**Original XML:** the original annotations also include `smoke`. XML and TXT therefore have different class coverage; the XML fire object count matches the corresponding TXT count for every image. XML files retain their original `<path>` values; those values refer to the source annotation environment and should not be used as download paths. Resolve images using `images/val/<filename>` instead.

For fire-only evaluation with the associated project, use `labels/val`. XML is supplied as the original additional annotation resource; do not assume that TXT includes smoke labels.

## Download

Install the Hugging Face client and download the dataset:

```bash
pip install --upgrade huggingface_hub
hf download yang1002/comprehensive-flame-dataset-validation --repo-type dataset --local-dir comprehensive-flame-dataset-validation
```

Or use Python:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yang1002/comprehensive-flame-dataset-validation",
    repo_type="dataset",
    local_dir="comprehensive-flame-dataset-validation",
)
```

For reproducible use, pass a specific commit SHA as `--revision` to the CLI or `revision=` to `snapshot_download`. All data files can be checked against `manifest.json` or `SHA256SUMS`.

## License and provenance

**No dataset license has been specified for this release.** Public availability does not itself grant a particular reuse license. Do not infer a data license from the project code, its dependencies, or the paper. Consult the dataset owner regarding reuse and any underlying source-specific terms.

The associated project describes data collected in experiments together with public fire-related data. This release preserves the existing validation files; it does not provide a verified per-image provenance map. See the project and paper for the experimental context.

## Related project and dataset

- Project code and dataset overview: [Small-Scale Fire Datasets](https://github.com/yang1002-life/small-fire_datasets).
- Related public release: [Binocular Flame Localization Dataset - Compact Subset](https://huggingface.co/datasets/yang1002/Binocular-Flame-Localization-Dataset), with 75 localization pairs and 20 calibration pairs (190 images in total), plus 150 Pascal VOC XML annotations for the localization images. It is a separate compact release, not an additional split of this validation dataset.
- Associated paper: [FireDet: A framework for precise detection and localization of small-scale fire](https://doi.org/10.1016/j.firesaf.2026.104737), *Fire Safety Journal*, Volume 162, Article 104737, 2026. See the citation below.

## Citation

```bibtex
@article{dai2026firedet,
  title   = {FireDet: A framework for precise detection and localization of small-scale fire},
  author  = {Dai, Jinyang and Cao, Qiuyang and Song, Xiaoning and Pan, Bence and Zhang, Qixing},
  journal = {Fire Safety Journal},
  volume  = {162},
  pages   = {104737},
  year    = {2026},
  doi     = {10.1016/j.firesaf.2026.104737}
}
```