jojomoi-meme/YOLO_fire_detection

YOLOv8s Fire & Smoke Detection

YOLOv8s Fire & Smoke Detection is a real-time computer vision model designed to detect active flames and smoke plumes in images and video streams. Fine-tuned on a dataset of nearly 30,000 annotated images, it offers a solid balance between lightweight deployment and detection accuracy.

Model Overview

• Architecture: YOLOv8s (Ultralytics)
• Task: Object Detection (object-detection)
• Input Resolution: 800x800 px
• Classes:
  • 0: Fire — Active flames
  • 1: Smoke — Smoke plumes

Performance Metrics

Evaluated on 3,363 test images using optimal inference settings (conf=0.333).

Metric	Score
mAP@50	62.55%
mAP@50-95	39.72%
Precision	64.95%
Recall	58.07%

---

Training Details

Parameter	Value
Base Model	YOLOv8s
Training Set Size	29,656 images
Image Size	800x800
Epochs	43
Batch Size	64
Optimizer	AdamW
Initial Learning Rate	0.001
Patience	10
Training Time	6h42

---

Quick Start & Usage

1. Installation

pip install ultralytics

Image Detection:

from ultralytics import YOLO

# Load the trained model
model = YOLO("path/to/best.pt")

# Run inference using the optimal confidence threshold
results = model.predict("image.jpg", conf=0.333)

# Process and display detections
for result in results:
    for box in result.boxes:
        cls_id = int(box.cls[0])
        conf = float(box.conf[0])
        label = model.names[cls_id]
        print(f"Detected: {label} ({conf:.2%})")

Video file:

from ultralytics import YOLO

model = YOLO("path/to/best.pt")
results = model.predict(source="video.mp4", conf=0.333, save=True)

Real Time detection:

from ultralytics import YOLO

model = YOLO("path/to/best.pt")
results = model.predict(source=0, conf=0.333, show=True)

Training Data:

Exemple output:

Intended Use Cases:

​Wildfire Early Warning: Drone and CCTV monitoring systems in forest areas. ​Smart Safety Monitoring: Automated alerts for industrial facilities, homes, and commercial buildings.

Limitations

​Low-Light Scenarios: Accuracy may drop in night-time scenes without adequate illumination. ​Environmental Distractions: Heavy fog, dust, or steam might occasionally trigger false positive smoke detections. ​Domain Adaptation: Optimal performance is achieved on visual settings closely aligned with the training distribution.

Dataset Information

​Source: Roboflow Universe Fire Detection Dataset (40,900 images). ​Train / Test Split: 29,656 images for training, 3,363 images for evaluation.