yang1002/Binocular-Flame-Localization-Dataset

Binocular Flame Localization Dataset - Compact Subset

This repository provides a compact subset of binocular flame-localization data from the original 20241227 collection, supplemented with corresponding calibration images from 20241227_origin/calibration/all/right. It is associated with the Small-Scale Fire Datasets project and the paper FireDet: A framework for precise detection and localization of small-scale fire, published in Fire Safety Journal, Volume 162, Article 104737, 2026. It is also associated with RNFD: An Adaptive RGB-NIR Heterogeneous Fusion System for Robust Fire Detection and Precise Localization, published in International Journal of Intelligent Systems, 2026, Article 7387105.

The release contains 270 original-resolution, lossless PNG images and 210 Pascal VOC XML detection annotations. Image0 and Image1 are RGB; Image2 contains grayscale near-infrared (NIR) images. These counts describe this compact release, not the full dataset totals reported by the associated papers. No training/test split is defined.

For a separate visible-light detection release, see the Comprehensive Flame Dataset - Validation Split, which contains 2,424 validation images with YOLO TXT and original XML annotations. That release is not a split of this binocular subset.

Dataset composition

Each of the 15 distance-group directories contains five same-numbered Image0 + Image1 RGB pairs. The first 12 groups also contain the five same-numbered, annotated Image2 NIR images. Selected Image2 images in 130.484, 140.201, and 150.271 have no XML annotations, so they are omitted. The localization portion has 210 images and 210 XML files.

The calibration portion contains 20 same-numbered RGB/NIR triplets, or 60 images: left contains Image0, the supplementary right directory contains Image1, and infra contains Image2. Calibration images have no detection XML files.

Distance-group directory	RGB pairs	Images	XML files
10.646	5	15	15
20.444	5	15	15
30.594	5	15	15
40.603	5	15	15
50.601	5	15	15
60.912	5	15	15
70.476	5	15	15
80.753	5	15	15
90.673	5	15	15
100.374	5	15	15
110.452	5	15	15
120.808	5	15	15
130.484	5	10	10
140.201	5	10	10
150.271	5	10	10
calibration/left	See total	20	0
calibration/right	See total	20	0
calibration/infra	See total	20	0
Total	75 localization pairs and 20 calibration triplets	270	210

All images are 5472 x 3648 pixels. RGB images have three channels; NIR images are single-channel grayscale. Distance-group directory names are preserved verbatim; their units have not been confirmed. Image0 and Image1 are original camera identifiers, and their physical left/right mapping for the localization images has not been confirmed. The original calibration directory name left is not additional evidence for the physical orientation of the localization cameras.

Sampling and lossless conversion

Within each distance group, complete Image0 + Image1 pairs are sorted by image number. Five pairs are selected using zero-based indices floor(i * (N - 1) / 4), for i = 0, ..., 4. In the first 12 groups, same-numbered Image2 images are included only where source XML annotations exist. For calibration, the sorted left images are sampled at indices floor(i * (N - 1) / 19), for i = 0, ..., 19; same-numbered images from the supplementary right directory and original calibration/infra directory form 20 triplets. Both rules include the first and last entries.

Sample identifiers from the 70 directory are already represented in 70.476, so they are not sampled twice. GA outputs, unrelated programs, caches and other processing outputs are excluded; build_subset.py is retained as the reproducible construction script. BMP-to-PNG conversion does not resize, crop or change the image mode. Each output was compared with its source using dimensions, mode and the SHA-256 of decoded pixels; all decoded pixels were verified identical.

File organization and annotations

10.646/ ... 120.808/      # Five Image0 / Image1 / Image2 triplets per group
130.484/ ... 150.271/     # Five Image0 / Image1 pairs per group
calibration/left/        # 20 Image0 PNGs
calibration/right/       # 20 same-numbered Image1 PNGs
calibration/infra/       # 20 same-numbered Image2 PNGs
labelimg/<distance_group>/  # XML annotations for the selected localization images
metadata.csv            # 270 rows of image metadata
manifest.jsonl          # Source, output, pixel and annotation verification records
summary.json            # Counts, sizes and verification results
SHA256SUMS.txt          # SHA-256 of all release files except this checksum file
build_subset.py         # Reproducible sampling and verification script
README.md               # Dataset card

Image numbers are unchanged; only the extension was converted from .bmp to .png. Each of the 210 localization images has a corresponding Pascal VOC XML annotation. XML filename values were updated to the PNG filename, and path values are relative to the dataset root. Classes, bounding-box coordinates and all other annotation content are unchanged. Resolve XML image paths relative to the dataset root.

metadata.csv is UTF-8 and contains these fields: file_name (relative image path), sample_id (group identifier prefixed by distance group or calibration), distance_group (original directory name; empty for calibration), camera_id, purpose (distance or calibration), annotation_file (relative XML path; empty for calibration), and modality (RGB or NIR). Images in the same numbered group share a sample_id. Read distance_group as a string.

Download

Install the Hugging Face client and download this compact subset:

pip install --upgrade huggingface_hub
hf download yang1002/Binocular-Flame-Localization-Dataset --repo-type dataset --local-dir Binocular-Flame-Localization-Dataset

Or use Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yang1002/Binocular-Flame-Localization-Dataset",
    repo_type="dataset",
    local_dir="Binocular-Flame-Localization-Dataset",
)

For reproducible downloads, pass a specific commit SHA as --revision to the CLI or revision= to snapshot_download. metadata.csv links each image through file_name; see the Hugging Face image dataset documentation for additional context. The release whitelist consists of the files listed in SHA256SUMS.txt, plus that checksum file itself; Windows thumbnail caches and other extra files are excluded. Hugging Face's repository metadata file .gitattributes is outside this release whitelist.

Reading example

import csv
from pathlib import Path
from PIL import Image

root = Path("Binocular-Flame-Localization-Dataset")
with (root / "metadata.csv").open(encoding="utf-8", newline="") as f:
    samples = list(csv.DictReader(f))
with Image.open(root / samples[0]["file_name"]) as image:
    print(samples[0]["sample_id"], image.size)

Verification and size

The release contains 270 images and 210 XML files. All 270 image conversions were verified to preserve decoded pixels, and all 210 annotations were verified unchanged except for filename and path updates. The 190 PNG images and 150 XML files from the first compact release retain their original hashes. These checks establish preservation during conversion; they do not independently establish annotation quality.

The selected source BMP images total 12.084 GiB; the PNG images total 4.422 GiB, or 36.59% of the original BMP size (a reduction of 63.41%). See summary.json for the release's total size, including documentation and checksum files but excluding the Hub-generated .gitattributes file.

manifest.jsonl records SHA-256 values for source images, PNG outputs and annotations. Records with source_collection=main use source paths relative to the original 20241227 collection; those with source_collection=calibration_right use source paths relative to the supplementary right directory. SHA256SUMS.txt checks the integrity of downloaded release files. On systems with sha256sum, run the following from the dataset root:

sha256sum --check SHA256SUMS.txt

Reproducible construction

Python 3.10+ and Pillow are required. The destination must be a new directory:

python build_subset.py --source <original-20241227-directory> --calibration-right-source <supplementary-right-directory> --destination <new-output-directory> --workers 3

If construction is interrupted, add --resume with the same output directory. Existing PNGs are still reverified against their source images; verification is not skipped. Reconstruction requires the original source collections, which are not included in this compact release.

Usage limitations, license and provenance

This is a small subset intended for demonstration, download and workflow validation. No training/test split has been defined. Calibration-board specifications, camera parameters and distance units have not been confirmed. The original annotations are retained without an additional claim about their quality.

No dataset license has been specified for this release. License and authorization terms remain to be supplied by the data owner. Public availability does not itself grant a particular reuse license; do not infer one from the project code, dependencies or paper. Consult the dataset owner about reuse and applicable source-specific terms.

Citation

If you use this release, cite the associated papers as applicable:

@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}
}

@article{dai2026rnfd,
  title   = {RNFD: An Adaptive RGB-NIR Heterogeneous Fusion System for Robust Fire Detection and Precise Localization},
  author  = {Dai, Jinyang and Cao, Qiuyang and Song, Xiaoning and Zhang, Qixing},
  journal = {International Journal of Intelligent Systems},
  year    = {2026},
  pages   = {7387105},
  doi     = {10.1155/int/7387105}
}

This repository is a compact localization subset; it does not contain all data used in either paper.