Evaluation of synthetic data impact on fire segmentation models performance

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Effective industrial fire detection is critical for safety.
• Deep neural networks offer superior fire detection but require extensive, high-quality datasets.
• Generating realistic fire datasets is challenging and resource-intensive.

Purpose of the Study:

• Introduce the SYN-FIRE dataset, a novel collection of simulated industrial fire images.
• Evaluate the impact of synthetic data on deep learning fire detection model performance.
• Assess the generalizability of models trained with synthetic data on real-world fire images.

Main Methods:

• Developed the SYN-FIRE dataset using NVIDIA Omniverse, comprising 2000 labeled synthetic fire images.
• Trained over 200 U-Net++ models using combinations of SYN-FIRE and public datasets.
• Conducted ablation studies involving data substitution and augmentation with synthetic data.
• Utilized GRAD-CAM saliency maps to analyze model generalization.

Main Results:

• Incorporating synthetic data improved Dice Scores by up to [Formula: see text] (FireBot) and [Formula: see text] (BowFire).
• Replacing real data with synthetic data generally enhanced performance, with minor exceptions.
• Models trained with synthetic data demonstrated improved generalization on challenging real-life fire images.

Conclusions:

• The SYN-FIRE dataset effectively enhances deep learning-based industrial fire detection.
• Synthetic data generation is a viable solution to overcome limitations of real-world data scarcity.
• The publicly available SYN-FIRE dataset facilitates future advancements in fire detection research.