DBA_Fire_Smoke-corrected

Fire and smoke detection — cleaned, de-duplicated, split and standardized version of the DBA-Fire dataset. 火焰与烟雾检测 —— DBA-Fire 数据集的清理、去重、划分与规整版。

English · 中文

---

English

What is this?

A cleaned, de-duplicated, split and standardized object-detection version of the DBA-Fire dataset — real-world fire and smoke detection imagery covering small indoor flames up to large-scale outdoor fires (city roads, buildings, supermarkets, parking lots). Reference paper: "DBA-YOLO: A Lightweight Fire Detection Algorithm with Structural Reparameterized Multiple Attention Mechanisms."

The original release ships as a flat, unsplit JPEGImages/ + txt/ pair (YOLO format, no configuration files, no split). It contains 57 groups of byte-identical duplicate images, mixes several source collections with inconsistent filenames (521 names contain spaces/parentheses), and shows heavy perceptual redundancy (UAV/video frames and re-encoded copies).

This release removes the exact duplicates, keeps near-identical images together on one side of a reproducible split, normalises the filenames, and ships ready-to-train YOLO TXT + Pascal VOC XML annotations with configuration files.

⚠️ Upstream license: none stated. The upstream README refers only to "[appropriate license terms]" and to "research purposes". See License & attribution — verify before commercial use.

Corrections vs. the official release

#	Correction	Detail
1	Removed 59 duplicate images	57 groups of images were byte-identical (md5; several share a UUID under two different names). The redundant copy of each group was removed.
2	Near-duplicates kept together	Many images are perceptually almost identical (consecutive video frames and re-encoded copies — captured by an 8×8-dHash ≤ 2 and 32×32 pixel-MSE ≤ 10 test). They are not deleted; instead each of the 3,014 groups (largest: 87 images; 1,028 images live in multi-image groups) is assigned entirely to one split, so near-identical content can never straddle train/val/test.
3	Reproducible train/val/test split	Upstream ships no split at all. This release provides a seeded, class-stratified 70 / 15 / 15 split (seed 42) over the near-duplicate groups.
4	Filename hygiene	521 filenames contained spaces and parentheses (e.g. img (74)) — hostile to scripts and cross-platform tooling. All names are normalised to the portable set [A-Za-z0-9._-].
5	Dual format + packaging	Upstream is YOLO TXT only. This release adds Pascal VOC XML from the same coordinates, plus classes.txt, data.yaml and per-split CSV tables.

Image pixels are never modified. All 3,846 images are byte-for-byte identical to the upstream files.

Scope — structural clean-up only. This is not a re-annotation. We de-duplicate, group and re-package the data; we do not re-verify individual boxes (missed fires, wrong class labels, box tightness).

Structural checks performed on the released set:

Check	Method	Result
Filename ↔ image ↔ label pairing	bidirectional	0 mismatches / orphans
Empty label files	file scan	0
Malformed rows / non-numeric fields	parse	0
Out-of-range boxes	xmin/ymin ≥ 0, xmax ≤ W, ymax ≤ H	0
Degenerate boxes	zero / negative width or height	0
Cross-split duplicate images	byte-identical md5 across splits	0
EXIF orientation anomalies	tag scan	0 (all orientation = 1)

Dataset at a glance

Property	Value
Images	3,846 (RGB .jpg, mixed resolutions)
Classes	2 — flame, smoke
Bounding boxes	7,988
Formats	Pascal VOC XML · YOLO TXT
Split	train 2,751 / val 542 / test 553 (seeded, class-stratified)

Boxes per class: flame 5,211 · smoke 2,777

Boxes per split: train 5,803 (flame 3,737 / smoke 2,066) · val 1,060 (768 / 292) · test 1,125 (706 / 419)

Structure

DBA_Fire_Smoke-corrected/
├── images/{train,val,test}/      # RGB .jpg (mixed resolutions)
├── labels/{train,val,test}/      # YOLO: cls cx cy w h (normalized)
├── Annotations/                  # Pascal VOC XML (flat, from the same coordinates)
├── JPEGImages/                   # copy of the images (VOC layout)
├── csv/
│   ├── train_labels.csv          # filename,width,height,class,xmin,ymin,xmax,ymax
│   ├── val_labels.csv
│   └── test_labels.csv
├── classes.txt                   # 2 class names, one per line
├── data.yaml                     # Ultralytics YOLO configuration
├── LICENSE
└── README.md

Every image has exactly one YOLO .txt and one VOC .xml; there are no empty label files.

Quick start (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

Citation

1. The original dataset / paper — please always cite this.

@misc{dba_fire_dataset,
  title        = {DBA-Fire: A Fire and Smoke Detection Dataset},
  author       = {{DBA-YOLO authors}},
  year         = {2025},
  howpublished = {Associated with "DBA-YOLO: A Lightweight Fire Detection Algorithm
                  with Structural Reparameterized Multiple Attention Mechanisms"},
  note         = {3905 images, YOLO-format fire and smoke annotations}
}

2. This corrected release — please cite it as well. It is not identical to the official release: duplicates were removed, a reproducible split was added, filenames were normalised and VOC annotations were generated, so a citation to the original alone does not describe this version.

@misc{dba_fire_smoke_corrected,
  author    = {KeenForgeAI},
  title     = {DBA_Fire_Smoke-corrected: a cleaned release of the DBA-Fire fire and smoke detection dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/DBA_Fire_Smoke-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from the DBA-Fire dataset.
               Upstream licence unstated.}
}

3. The annotation tool (optional).

@software{keenforge,
  author    = {KeenForgeAI},
  title     = {KeenForge: a local-first, offline image annotation and model-training desktop tool},
  year      = {2026},
  publisher = {KeenForgeAI},
  url       = {https://github.com/KeenForgeAI/KeenForge},
  note      = {MIT licensed. Developed by Lu Gan and Sam Li.}
}

License & attribution

The upstream DBA-Fire dataset states no license. Its README ships no LICENSE file, refers only to "[appropriate license terms]" and states the data is available "for research purposes". The copyright status of the original images and annotations is therefore undetermined — verify it before any commercial use. See LICENSE in this repository.

Our modifications (de-duplication, grouping, split, filename normalisation, format conversion, documentation) are released under CC BY 4.0. Image pixels are unmodified and remain subject to the upstream terms.

Provenance

• Source: DBA-Fire dataset (YOLO format), retrieved 2026-10-13.
• Upstream: 3,905 images / 8,087 boxes (flame 5,310 · smoke 2,777); no split; no config files.
• Verified: 3,846 images (7,988 boxes, byte-identical); 59 exact duplicates removed; 3,014 near-duplicate groups; 0 cross-split leakage, 0 invalid boxes.

---

中文

这是什么？

DBA-Fire 数据集的清理、去重、划分与标准化目标检测版——真实场景的火焰与烟雾检测图像， 覆盖从小型室内火焰到城市道路、楼宇、超市、停车场等大型户外火灾。 参考论文《DBA-YOLO: A Lightweight Fire Detection Algorithm with Structural Reparameterized Multiple Attention Mechanisms》。

原始版本以扁平的、无划分的 JPEGImages/ + txt/（YOLO 格式，无配置文件、无划分）发布， 包含 57 组逐字节完全重复的图片，混杂了多个来源且文件名不一致（521 个文件名含空格/括号）， 并存在大量感知重复（视频/连拍帧与重编码副本）。

本版本删除完全重复、把近重复图整组固定到同一划分、规整文件名，并提供开箱即用的 YOLO TXT 与 Pascal VOC XML 双格式标注及配置文件。

⚠️ 上游未声明许可证。 上游 README 仅提到 "[appropriate license terms]" 与 "research purposes"。详见 许可证与署名——商用前请自行确认。

相对官方版的修正

#	修正	说明
1	删除 59 张完全重复图	有 57 组图片逐字节完全相同（md5 相同；其中多组是同一 UUID 的两个不同文件名）。本版删除每组中多余的副本。
2	近重复图固定同侧	大量图片在感知上几乎相同（连续视频帧与重编码副本——以 8×8 dHash ≤ 2 且 32×32 像素 MSE ≤ 10 判定）。它们不删除，而是把 3,014 个分组（最大一组 87 张；共 1,028 张位于多图分组中）整组分配到一个划分，确保近似内容绝不跨越 train/val/test。
3	可复现的训练/验证/测试划分	上游完全没有划分。本版基于近重复分组，提供固定随机种子、按类别分层的 70 / 15 / 15 划分（随机种子 42）。
4	文件名净化	521 个文件名含空格与括号（如 img (74)），对脚本与跨平台工具均不友好。现全部规整为跨平台安全的 [A-Za-z0-9._-]。
5	双格式 + 规范化打包	上游仅有 YOLO TXT；本版由同一坐标生成 Pascal VOC XML，并补充 classes.txt、data.yaml 与按划分的 CSV 表格。

图像像素从不修改——全部 3,846 张图与上游文件逐字节一致。

本版范围——仅结构化清理。 这是结构性清理，不是重新标注。我们只做去重、归组、重新打包； 不重新核对单个标注框（漏标、类别标错、框松紧）。

已执行的结构化检查（针对发布的集合）：

检查	方法	结果
文件名↔图片↔标注 配对	双向	0 错配 / 孤儿
空标签文件	文件扫描	0
坏行 / 非数字字段	解析	0
越界框	xmin/ymin ≥ 0，xmax ≤ W，ymax ≤ H	0
退化框	宽或高为零/负	0
跨划分重复图	跨划分逐字节 md5	0
EXIF 方向异常	标签扫描	0（方向全部为 1）

数据集概览

属性	值
图片	3,846（RGB .jpg，分辨率不一）
类别	2 —— flame、smoke
标注框	7,988
格式	Pascal VOC XML · YOLO TXT
划分	train 2,751 / val 542 / test 553（固定种子、类别分层）

各类框数： flame 5,211 · smoke 2,777

各划分框数： train 5,803（flame 3,737 / smoke 2,066）· val 1,060（768 / 292）· test 1,125（706 / 419）

目录结构

DBA_Fire_Smoke-corrected/
├── images/{train,val,test}/      # RGB .jpg（分辨率不一）
├── labels/{train,val,test}/      # YOLO：cls cx cy w h（归一化）
├── Annotations/                  # Pascal VOC XML（扁平，来自同一坐标）
├── JPEGImages/                   # 图片副本（VOC 布局）
├── csv/
│   ├── train_labels.csv          # filename,width,height,class,xmin,ymin,xmax,ymax
│   ├── val_labels.csv
│   └── test_labels.csv
├── classes.txt                   # 2 个类别名称，每行一个
├── data.yaml                     # Ultralytics YOLO 配置文件
├── LICENSE
└── README.md

每张图片都有且仅有一个 YOLO .txt 与一个 VOC .xml，不存在空标签文件。

快速开始 (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

引用

使用本数据集请同时引用原始工作与本修正版：

1. 原始数据集 / 论文（请务必引用） —— 见上方英文部分 dba_fire_dataset。

2. 本修正版（请一并引用） —— 本版与官方发布并不相同：删除了重复图、补充了可复现划分、 规整了文件名并生成了 VOC 标注，因此只引用原数据集无法描述本版本。见上方英文部分 dba_fire_smoke_corrected。

3. 标注工具（可选） —— 见上方英文部分 keenforge。

许可证与署名

上游 DBA-Fire 数据集未声明许可证。 其 README 无 LICENSE 文件，仅提到 "[appropriate license terms]" 并说明数据 "for research purposes"。因此原始图像与标注的版权状态未定—— 商用前请自行确认。详见本仓库 LICENSE。

我们对数据集的修改部分（去重、归组、划分、文件名净化、格式转换、文档）以 CC BY 4.0 发布。 图像像素未修改，仍受上游条款约束。

来源

• 来源：DBA-Fire 数据集（YOLO 格式），获取于 2026-10-13
• 上游：3,905 图 / 8,087 框（flame 5,310 · smoke 2,777）；无划分；无配置文件
• 已校验：3,846 图（7,988 框，逐字节一致）；删除 59 张完全重复；3,014 个近重复分组；0 跨划分泄漏、0 非法框