[ICIP 2025 SATELLITE WORKSHOP] UNREALFIRE: A SYNTHETIC DATASET CREATION PIPELINE FOR ANNOTATED FIRE IMAGERY IN UNREAL ENGINE

Paper / Dataset

This repository contains:

• Modified AirSim files with our Particle segmentation camera presented in the paper
• Some extra tools for mask re-coloring and the dilation used in the paper.
• Link to download the AUW (Auth-Unreal-Wildfire) fire segmentation dataset we generated through this pipeline.

Prerequisites

• You need to have Unreal Engine Version 5.2 installed.
• Download AirSim/Colloseum from the original Repository.
• We used the M5 VFX Vol2. Fire and Flames(Niagara) for fire assets and multiple sources for vegetation assets.

Instructions

• (Optional) Use World Machine to create a custom landscape and/or Blender for custom 3D assets.
• Create a new project in UE and add the actor C++ files from the Source folder to the project's source folder.
• Rewrite the project name in the actor files accordingly.
• Follow the default instructions from the original AirSim/Colloseum repository, all the way up to the installation step.
• Replace the downloaded AirSim/Source/PIPCamera.cpp and add AirSim/Content/HUDAssets from this repository.
• Continue with the AirSim installation.
• Add the Procedural Content Generation (PCG) graph to a PCG volume to generate actors.
• To add the particle segmentation camera, the ImageType is 10. { "CameraName": "0", "ImageType": 10, "PixelsAsFloat": false, "Compress": true }
• You can either record using the built-in recorder function from the editor or through the Python API.

Dataset

• After obtaining the RGB images and Segmentation masks, place the helper/binary.py file inside the masks folder to the masks binary.
• (Optional) Run the helper/mask\_wide.py to smooth and dialite the generated masks.

AUW Dataset

In our paper, we gathered 1700 images through the UnrealFire pipeline and created AUW. The dataset is open-access and can be found here. Some Benchmark results can be found in Table 1 below. For a comprehensive presentation of UnrealFire and AUW, please refer to the paper. The RGB color vectors of the masks are [255,0,0] for fire, and [170,40,120] for background. There is a premade split (75% training - 25% validation), but you are encouraged to try other splits.

Table 1: Results on the test set of Corsican

The +X% represents adding this X percentage of training images from the Corsican training set.
75% represents the whole Corsican training set.

Training Dataset	Fire IoU	mIoU
Corsican	86.95	91.60
AUW	45.07	65.86
AUW + 1%	78.17	86.06
AUW + 2%	81.33	87.99
AUW + 5%	83.80	89.49
AUW + 10%	87.52	91.98
AUW + 25%	87.22	91.82
AUW + 50%	87.80	92.14
AUW + 75%	89.35	93.18

Citation

If you find our work useful, please consider citing and starring:

@inproceedings{spatharis2025unrealfire,
  title={Unrealfire: A Synthetic Dataset Creation Pipeline for Annotated Fire Imagery in Unreal Engine},
  author={Spatharis, Evangelos and Papaioannidis, Christos and Mygdalis, Vasileios and Pitas, Ioannis},
  booktitle={2025 IEEE International Conference on Image Processing Workshops (ICIPW)},
  pages={610--615},
  year={2025},
  organization={IEEE}
}