Diffusion-Detect: Synthetic and Real Data Towards a Robust Fire and Smoke Detection

Video surveillance systems enhance security and monitoring in public spaces and critical infrastructure, but their effectiveness relies on diverse, high-quality data for training and evaluation. In this context, we explore the potential of synthetic data and generative AI in video surveillance to address events that are rare and difficult to gather in the real world. We focus on fire and smoke detection by proposing diffusion detection, a methodology for generating rare synthetic fire and smoke events. Such synthetic data will be used to complement real data to train and assess a YOLO-based fire and smoke detector. The experimental analysis we report, based on two real datasets and two synthetic datasets we generate, is devoted to a critical analysis of the potential of injecting synthetically generated data in training and testing procedures.