Datasetpublic

baizhanquan/FireDetectionDataset-flame-forest-flameye-wildfire

FlamEye — Wildfire Detection Dataset A merged, deduplicated, and augmented dataset for real-time wildfire detection (fire and smoke) from CCTV/surveillance cameras. Built to train YOLOv8m for early-stage fire detection. Classes ID Name 0 fire 1 smoke Dataset Statistics Split Images Train ~10,929 Validation ~3,000 Test ~1,500 Sources Dataset Source Notes D-Fire Kaggle Class IDs remapped:… See the full description on the dataset page: https://huggingface.co/datasets/baizhanquan/FireDetectionDataset-flame-forest-flameye-wildfire.

sourceHugging Facecc-by-4.0updated 1mo agoView on Hugging Face

2likes364downloads

Dataset Card

FlamEye — Wildfire Detection Dataset

A merged, deduplicated, and augmented dataset for real-time wildfire detection (fire and smoke) from CCTV/surveillance cameras. Built to train YOLOv8m for early-stage fire detection.

Classes

ID	Name
0	fire
1	smoke

Dataset Statistics

Split	Images
Train	~10,929
Validation	~3,000
Test	~1,500

Sources

Dataset	Source	Notes
D-Fire	Kaggle	Class IDs remapped: original smoke=0/fire=1 → fire=0/smoke=1
Pyro-SDIS	HuggingFace	Forest fire surveillance footage
AI for Mankind	Manual upload	VOC XML converted to YOLO format
Catargiu 2024	GitHub	Additional fire/smoke samples

Processing Pipeline

1.Normalize — All sources remapped to unified class IDs (fire=0, smoke=1)
2.Deduplicate — pHash deduplication with threshold=8 across all sources
3.Split — 70/20/10 stratified train/val/test split
4.Augment — 2× augmentation on train split, 4× for fire-only images (underrepresented class)

Format

Each example contains:

—image — PIL Image (RGB)
—image_id — filename stem
—split — train, validation, or test
—source — original dataset name (dfire, pyro_sdis, aiformankind, catargiu)
—annotations — list of {class_id, class_name, x_center, y_center, width, height} (YOLO normalized)
—has_fire — bool
—has_smoke — bool

Usage

python

from datasets import load_dataset

ds = load_dataset("Hajorda/flameye-wildfire-detection")
train = ds["train"]

# Filter fire-only images
fire_images = train.filter(lambda x: x["has_fire"] and not x["has_smoke"])

—Model: Hajorda/flameye-yolov8m (coming soon)
—Training target: mAP@50 > 0.88, Recall > 0.90

License

CC BY 4.0 — Please credit the original source datasets if you use this.