High Quality Fire Smoke Dataset: A Benchmark for Fire and Smoke Detection

In this paper, we present the High Quality Fire Smoke Dataset(HQFSD), a new comprehensive fire and smoke dataset tailored for training and evaluating fire detection algorithms. It currently comprises 12,166 meticulously selected images sourced from over 250 real-fire video clips available on the Internet. These images encompass a wide range of fire scenarios, including diverse geographical and environmental conditions, varying lighting, background, burning objects, processes and states, and levels of occlusion. Each image is annotated with classification labels and bounding boxes. Based on the HQFSD, we systematically study and benchmark 14 representative deep learning methods. We also present the testing results to researchers utilizing the dataset and provide an overall analysis of these 14 baseline methods. Furthermore, we compare HQFSD with another fire dataset, demonstrating its robust generalization capability. It is anticipated that the HQFSD will significantly contribute to the advancement of fire detection research and applications.