---
license: agpl-3.0
library_name: ultralytics
pipeline_tag: object-detection
tags:
- yolo
- yolo11
- fire-detection
- computer-vision
- realtime
---

# YOLOv11n Fire Detector (ProFSAM)

**Paper:** https://arxiv.org/abs/2510.21782  
**Code:** https://github.com/UEmmanuel5/ProFSAM  
**Weights:** `Fire_best.pt`  

## Intended use
Bounding-box detection of fire to prompt SAM2/MobileSAM/TinySAM in the ProFSAM pipeline.

## Training data
FASDD subset: classes **fire** and **neither_firenorsmoke** only. Total images used: **51,749** (12,550 fire, 39,199 neither_firenorsmoke).

## Training setup (summary)
PyTorch 2.0, CUDA 12.4, main GPU: GTX 1050 Ti (4 GB).  
Ultralytics YOLOv11n initialized then trained 100 epochs.

### Script used
```python
# train_yolo11n.py
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
from ultralytics import YOLO
import torch
torch.backends.cudnn.benchmark = True

model = YOLO("path/to/yolo11n.pt")
train_results = model.train(
    data="path/to/FASDD_CV_Fire/data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    optimizer="AdamW",
    lr0=1e-4,
    lrf=0.01,
    dropout=0.15,
    weight_decay=5e-4,
    device=0,
    val=False,
    save=True,
    plots=False
)
````

## Detector metrics (FASDD fire-only subset)

| P     | R     | mAP@0.5 | mAP@0.5:0.95 |
| ----- | ----- | ------- | ------------ |
| 0.799 | 0.697 | 0.797   | 0.520        |


## Test data
If you do not have test images, I placed 4 test images from the khan dataset to be used during your testing phase.