yang1002/RGB-NIR-Flame-Dataset-Validation

RGB-NIR Flame Dataset - Compressed Validation Split

This repository provides only the existing 960 x 640 compressed validation split of the RGB-NIR flame detection dataset associated with the RGB_NIR project and the published RNFD paper.

It contains 3,171 paired RGB-NIR samples (6,342 JPEG images) with per-modality YOLO TXT and Pascal VOC XML annotations. Images and annotations are preserved byte for byte from the existing compressed validation files. No further image compression, annotation conversion, resampling or split generation was performed for this release.

No training data, original-resolution images or independent localization dataset are included. The source directories retain the name val; the paper describes a 3,171-pair detection test split. This release preserves the existing val names and membership rather than renaming or redefining that split. The counts below describe this release, not the complete resources described in the paper.

Release contents

Content	Count / value
Paired RGB-NIR samples	3,171
RGB JPEG images	3,171
NIR JPEG images	3,171
Image dimensions	960 x 640
YOLO TXT files	6,342 (3,171 per modality)
Pascal VOC XML files	6,342 (3,171 per modality)
Image and annotation files, total	19,026
RGB fire bounding boxes	1,006
NIR fire bounding boxes	991
Empty RGB YOLO label files	2,209
Empty NIR YOLO label files	2,227

The image and annotation payload occupies 1,011,421,171 bytes (1.011 GB, decimal). Additional pairing, manifest and documentation files are small; release_info.json records data counts and byte sizes.

images/val_rgb/          # 3,171 RGB JPEG images (Image0_<pair_id>.jpg)
images/val_infra/        # 3,171 NIR JPEG images (Image1_<pair_id>.jpg)
labels/val_rgb/          # RGB YOLO TXT labels
labels/val_infra/        # NIR YOLO TXT labels
annotations/val_rgb/     # Original RGB Pascal VOC XML annotations
annotations/val_infra/   # Original NIR Pascal VOC XML annotations
pairs.csv               # 3,171 rows of paired image and annotation paths
manifest.jsonl          # Paths, kinds, modalities, byte sizes and SHA-256 of data files
release_info.json       # Verified release counts and data byte size
SHA256SUMS.txt          # Checksums of all release files except this checksum file
README.md

Training files, original-resolution images, localization data, caches, operating-system thumbnails, filtered/oversampled experiments and duplicate copies are excluded. Hugging Face's repository metadata file .gitattributes is outside the release checksum inventory.

Pairing and annotation formats

Each sample has one RGB image named Image0_<pair_id>.jpg and one NIR image named Image1_<pair_id>.jpg. Remove only the respective Image0_ or Image1_ prefix and the extension to obtain the pair identifier. Treat identifiers as strings, preserving minus signs and leading zeros. Their physical camera left/right orientation is not inferred here. Images with the same pair identifier belong together, but their annotations are modality-specific and should not be assumed identical.

pairs.csv contains pair_id, rgb_image, nir_image, rgb_yolo, nir_yolo, rgb_xml and nir_xml. All file paths are relative to the dataset root. The CSV has one row per pair, not one row per image.

YOLO TXT: every nonempty line is 0 x_center y_center width height, with coordinates normalized to image dimensions. Class 0 is fire. An empty TXT file is retained and indicates no annotated fire boxes for that image; do not discard the image or its paired modality.

Pascal VOC XML: annotations use the class name fire and pixel-space xmin, ymin, xmax, ymax coordinates for the 960 x 640 image. XML bytes are unchanged; only their release directory is organized under annotations/val_rgb or annotations/val_infra. Resolve images from pairs.csv or images/val_<modality>/<filename> rather than interpreting source XML folder metadata as download paths.

Download

Install the Hugging Face client and download the complete compressed validation release:

pip install --upgrade huggingface_hub
hf download yang1002/RGB-NIR-Flame-Dataset-Validation --repo-type dataset --local-dir RGB-NIR-Flame-Dataset-Validation

Or use Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yang1002/RGB-NIR-Flame-Dataset-Validation",
    repo_type="dataset",
    local_dir="RGB-NIR-Flame-Dataset-Validation",
)

For a reproducible download, specify a commit SHA using --revision in the CLI or revision= in Python. On systems with sha256sum, run sha256sum --check SHA256SUMS.txt from the downloaded dataset root.

Reading paired samples

import csv
from pathlib import Path
from PIL import Image

root = Path("RGB-NIR-Flame-Dataset-Validation")
with (root / "pairs.csv").open(encoding="utf-8", newline="") as handle:
    pairs = list(csv.DictReader(handle))  # Keeps pair_id as a string.

sample = pairs[0]
with Image.open(root / sample["rgb_image"]) as rgb, Image.open(root / sample["nir_image"]) as nir:
    print(sample["pair_id"], rgb.size, nir.size)

Verification and limitations

All 6,342 images were fully decoded and checked for 960 x 640 dimensions. Both modalities have matching sample identifiers and corresponding TXT/XML files. All annotations were parsed and checked for the expected fire class, finite coordinates, positive box dimensions and image bounds. Per-image YOLO/XML object counts were compared, and their box coordinates checked within one pixel to allow integer XML rounding. Source and staged bytes were compared by SHA-256; manifest.jsonl records each of the 19,026 data files. These checks establish file preservation and structural consistency, not an independent assessment of annotation quality.

This compressed validation release is intended for evaluation, comparison and workflow reproduction. It does not provide the original-resolution training data, full calibration resources, model code or the paper's separate localization dataset. RGB and NIR annotations may differ because the two modalities are annotated separately.

No dataset license has been specified for this release. Public availability does not itself grant a particular reuse license. Do not infer a dataset license from the paper, project code or dependencies; contact the dataset owner about applicable reuse terms.

Associated publication

RNFD: An Adaptive RGB-NIR Heterogeneous Fusion System for Robust Fire Detection and Precise Localization

• Authors: Jinyang Dai, Qiuyang Cao, Xiaoning Song, and Qixing Zhang
• Journal: International Journal of Intelligent Systems, Volume 2026, Issue 1, Article 7387105
• First published online: July 30, 2026
• DOI: 10.1155/int/7387105
• Project: yang1002-life/RGB_NIR

RNFD uses complementary RGB and NIR imagery for flame perception, combining the RNAFNet detector and adaptive feature fusion with camera-parameter optimization, flame segmentation and binocular stereo matching.

Full dataset scale reported in the paper

Dataset / split	RGB-NIR image pairs
Detection dataset, total	17,977
Detection training split	14,806
Detection test split	3,171
Independent localization dataset	1,395

These figures describe the paper's resources. Only the existing compressed validation files described above are included in this repository.

Citation

@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},
  volume  = {2026},
  number  = {1},
  eid     = {7387105},
  year    = {2026},
  doi     = {10.1155/int/7387105},
  url     = {https://doi.org/10.1155/int/7387105}
}