FireDetection

Real-time flame and smoke detection with a fine-tuned YOLO11n model, plus an INT8 OpenVINO export for lighter CPU inference.

Model and data

• Base model: Ultralytics yolo11n.pt, fine-tuned for object detection.
• Dataset: referenced as data/dataset.yaml, which is not committed. The dataset's source and license are currently undocumented.

Training config

From train/args.yaml: 100 epochs, batch 16, image size 640. The training notebook is main.ipynb.

Results

Final epoch (epoch 100) metrics on the validation split, from train/results.csv:

Precision	Recall	mAP50	mAP50-95
0.935	0.904	0.945	0.651

INT8 OpenVINO export

train/weights/best_int8_openvino_model/ holds the INT8 export. The Ultralytics benchmark recorded in benchmarks.log (imgsz 640, same dataset):

Format	Size	mAP50-95	Inference	FPS
OpenVINO INT8	3.2 MB	0.6367	13.58 ms/img	73.61

The hardware used for training and benchmarking was not recorded.

Run it

pip install -r requirements.txt   # ultralytics; uses the PyTorch cu128 index
python test.py                    # Ultralytics inference app on train/weights/best.pt

Gaps

Dataset provenance, hardware, and a license still need to be documented.