FlameVideo

We collected nearly 20,000 flame images from multiple open-source platforms, and after rigorous screening, we finally retained more than 15,000 high-quality images that meet the resolution requirements. These images are sourced from the website https://aistudio.baidu.com/datasetdetail/281895/0.To fully reflect the different characteristics of flames, we classified these images into three categories based on flame scale and visual features: large fire, small fire, and smoke, ensuring that the dataset covers a rich variety of flame scenarios. All images have undergone unified normalization processing to meet the standardization requirements for deep learning model inputs.To simulate the dynamic changes of flames, we adopted a frame sequence method, combining every three consecutive images to generate a short video clip. Through this approach, we created the FlameVideo dataset containing 5,209 video samples of three flame conditions. The video resolution is uniformly set to 720*1280 with the mp4 format. Given that complex video backgrounds and different flame viewing angles can affect the accuracy of abnormal flame recognition, the research team developed a dataset with various complex backgrounds. The dataset is divided into a training set and a test set at a ratio of approximately 2:1, which are used for model training and performance evaluation respectively.