Evaluation of synthetic data impact on fire segmentation models performance

Timely fire detection in industrial environments is crucial to safeguarding people and property. Deep neural networks have shown effectiveness in fire detection over traditional methods. However, they require high-quality datasets, which are costly and time-intensive to gather. To overcome this issue, we created the SYN-FIRE dataset, consisting of 2000 labeled images of simulated indoor industrial fires using NVIDIA Omniverse. By using U-Net++ as the baseline, this study explores the impact of the new SYN-FIRE dataset on models' performance when combined with four publicly available datasets. Two ablation studies were conducted: one replacing portions of real data from publicly available datasets with synthetic data and the other adding various amounts of synthetic data. With over 200 models trained across three resolutions, the results indicate that incorporating additional synthetic data improved DiceScore by [Formula: see text] to [Formula: see text] (FireBot and BowFire datasets, respectively) while substituting real data with synthetic data generally enhanced performance but with exceptions. Furthermore, tests on challenging real-life fire images confirmed that synthetic data boosts model generalization, supported by GRAD-CAM saliency maps. Finally, we provide key takeaways that point out the main findings of our research. The SYN-FIRE dataset is publicly available to encourage further research in fire detection and prevention.