---
license: agpl-3.0
library_name: ultralytics
datasets:
  - sreeharivp23/fire-segmentation-dataset
pipeline_tag: image-segmentation
tags:
  - yolo
  - yolo11
  - ultralytics
  - instance-segmentation
  - fire-detection
  - safety
  - pytorch
---

# Fire Segmentation — YOLO11n-seg

A YOLO11n instance-segmentation model fine-tuned to detect **fire** with
pixel-accurate masks. Unlike bounding-box fire detectors, it reports the exact
flame region, which makes it possible to measure how much of a frame is
burning and to draw precise overlays for monitoring/alerting systems.

![example](example_output.jpg)

## Usage

```python
from ultralytics import YOLO

model = YOLO("fire_seg_yolo11n.pt")

# image, video file, folder, URL or webcam index all work
results = model.predict("fire_video.mp4", conf=0.35)

for r in results:
    if r.masks is not None:
        print(f"fire instances: {len(r.masks)}")
    r.show()  # or r.save()
```

Or from the CLI:

```bash
yolo segment predict model=fire_seg_yolo11n.pt source=fire_video.mp4 conf=0.35
```

## Model details

| | |
|---|---|
| Base model | `yolo11n-seg.pt` (Ultralytics YOLO11 nano, segmentation) |
| Task | Instance segmentation, 1 class: `fire` |
| Input size | 640 × 640 |
| Epochs | 60 |
| Mask mAP50 | 0.54 |
| Mask mAP50-95 | 0.29 |
| Box mAP50 | 0.55 |

![training curves](training_curves.png)

## Training data

The full auto-labelled dataset is published at
[sreeharivp23/fire-segmentation-dataset](https://huggingface.co/datasets/sreeharivp23/fire-segmentation-dataset):
~1,100 fire images from the
[CAIR Fire-Detection-Image-Dataset](https://github.com/cair/Fire-Detection-Image-Dataset)
and the
[DeepQuestAI Fire-Smoke-Dataset](https://github.com/DeepQuestAI/Fire-Smoke-Dataset),
plus 250 no-fire negatives.

Segmentation labels were **machine-generated**: a pretrained YOLOv8
fire/smoke bounding-box detector proposed fire regions, SAM 2.1 (base)
converted each box into a pixel mask, and masks were polygonised into
YOLO-seg format. An HSV flame-colour heuristic covered images the detector
missed. Validation metrics above are measured against these auto-generated
labels, not human annotations.

## Limitations

- Labels are machine-generated; expect some mask noise, especially around
  smoke/glow boundaries and small or occluded flames.
- Trained mostly on visible orange/yellow flames; performance on blue flames,
  night-vision or thermal imagery is untested.
- **Not a certified fire-safety system.** Do not use as a substitute for
  smoke/fire alarms.

## License

Fine-tuned from Ultralytics YOLO11, which is released under **AGPL-3.0**;
these weights inherit that license.