Early Fire & Smoke Detection

Object detection system for spotting small flames and early-stage smoke in indoor security camera footage. The goal is early detection, before a fire has a chance to spread, using two independently implemented and compared approaches.

Approaches

Both models share the same dataset, preprocessing pipeline (640x640 letterbox), and evaluation module, so the comparison is apples-to-apples:

1. EYT-Net — a single-stage detector built from scratch in PyTorch: custom CNN backbone, two-scale detection head (P4/P5), IoU-based k-means anchors, CIoU box loss + BCE objectness/classification loss, and a full training loop with logging, checkpointing, and early stopping.
2. Faster R-CNN — a two-stage detector using a torchvision ResNet-50-FPN-v2 backbone pretrained on COCO, fine-tuned for the fire / smoke classes with a staged unfreezing schedule and a short hyperparameter search (learning rate, backbone layers, LR schedule).

Results (test set, 1,300 images)

Model	mAP@0.5	mAP@0.5:0.95	Precision	Recall	F2	Speed (RTX 4060)
Faster R-CNN	0.931	0.589	0.877	0.918	0.909	~95.7 ms (10.4 FPS)
EYT-Net	0.365	0.170	0.413	0.453	0.444	~30.6 ms (32.6 FPS)

The pretrained two-stage model is clearly more accurate, especially on the harder smoke class. The from-scratch model is roughly 3x faster and much smaller (~7.2M vs ~43M parameters), trading accuracy for speed and independence from pretrained weights.

Dataset

Home Fire Dataset — 6,500 images, 2 classes (fire, smoke), split 3,900 / 1,300 / 1,300 (train/val/test). Downloaded automatically via kagglehub in notebooks/01_setup.ipynb.

Setup

Managed with uv.

uv sync
.venv\Scripts\activate

For GPU support, match your CUDA version in pyproject.toml ([[tool.uv.index]] URL, e.g. cu128/cu130), then re-run uv lock && uv sync. Verify with:

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

Project structure

data/          datasets (train/val/test images + YOLO-format labels)
eytnet/        EYT-Net model, dataset, loss, postprocessing, metrics
helpers/       shared preprocessing, augmentation, viz, training utils
notebooks/     data analysis, training, tuning, evaluation, demo
models/        checkpoints and run configs

Notebooks run roughly in order: data setup/EDA → preprocessing & augmentation → anchor analysis → Faster R-CNN fine-tuning & tuning → EYT-Net training → comparison, error analysis, and demo.