Fire Detection – Mask R-CNN Pipeline

Overview

This folder contains the full pipeline for training a Mask R-CNN fire instance segmentation model using synthetic data auto-generated from FDS simulations.

Pipeline Steps

Step 1 – Extract Frames (Smokeview)

Renders two image sequences per FDS scenario:

• RGB frames – realistic smoke/fire scene
• Mask frames – white fire isosurface (HRRPUV) on black background

python extract_frames.py
# Or test with a single scenario:
python extract_frames.py --scenario A1_small_room_opening_0

Step 2 – Prepare Dataset

Binarizes the masks and creates a manifest.csv with train/val/test splits.

python prepare_dataset.py

Step 3 – Train Mask R-CNN

Uses transfer learning from ResNet-50-FPN (ImageNet weights).

python train_maskrcnn.py --manifest Dataset/manifest.csv --epochs 25

Directory Layout

Fire Detection/
  Dataset/
    images/         ← RGB frames from Smokeview
    masks/          ← Raw HRRPUV masks from Smokeview
    masks_binary/   ← Binarized masks (generated by prepare_dataset.py)
    manifest.csv    ← Generated by prepare_dataset.py
  checkpoints/      ← Saved model weights (.pt)
  logs/             ← training_log.csv
  extract_frames.py
  prepare_dataset.py
  train_maskrcnn.py

Requirements

torch>=2.0
torchvision>=0.15
Pillow
numpy