sreeharivp23/fire-segmentation-dataset

Fire Segmentation Dataset (YOLO-seg format)

Instance-segmentation dataset for fire detection: 1,348 images with polygon mask labels in Ultralytics YOLO segmentation format. Built to train sreeharivp23/fire-segmentation-yolo11n.

Contents

Split	Images
train	1,146
val	202

• 1,098 fire images with one or more fire polygon instances
• 250 negatives (no fire) with empty label files

fire_seg/
├── data.yaml            # Ultralytics dataset config (1 class: fire)
├── images/{train,val}/  # .jpg / .png images
└── labels/{train,val}/  # YOLO-seg .txt labels: "0 x1 y1 x2 y2 ..." (normalised)

Usage

from huggingface_hub import snapshot_download
from ultralytics import YOLO

path = snapshot_download("sreeharivp23/fire-segmentation-dataset", repo_type="dataset")

model = YOLO("yolo11n-seg.pt")
model.train(data=f"{path}/data.yaml", epochs=60, imgsz=640)

Note: data.yaml contains a path: entry pointing at the original build location — update it to the downloaded folder (or delete the line and pass absolute paths) before training.

How the labels were made

Labels are machine-generated, not human-annotated:

1. A pretrained YOLOv8 fire/smoke bounding-box detector proposed fire regions.
2. SAM 2.1 (base) converted each box into a pixel mask.
3. Masks were polygonised (OpenCV contours + Douglas-Peucker) into YOLO-seg polygons.
4. Images the detector missed fell back to an HSV flame-colour heuristic (219 images); 12 images yielded nothing and were dropped.

Expect some label noise around smoke/glow boundaries and small flames. A YOLO11n-seg fine-tuned on this data reaches mask mAP50 ≈ 0.54 on the val split.

Image sources & licensing

Images originate from two public research datasets; this repackaging adds machine-generated segmentation labels:

• CAIR Fire-Detection-Image-Dataset
• DeepQuestAI Fire-Smoke-Dataset

Check the upstream repositories for their exact terms before commercial use.