CAM-based non-local attention network for weakly supervised fire detection

Many available object detectors are already used in fire detection, such as Faster RCNN, SSD, YOLO, etc., to localize the fire in images. Although these approaches perform well, they require object-level annotations for training, which are manually labeled and very expensive. In this paper, we propose a method based on the Class Activation Map (CAM) and non-local attention to explore the Weakly Supervised Fire Detection (WSFD) given only image-level annotations. Specifically, we first train a deep neural network with non-local attention as the classifier for identifying fire and non-fire images. Then, we use the classifier to create a CAM for every fire image in the inference stage and finally generate a corresponding bounding box according to each connected domain of the CAM. To evaluate the availability of our method, a benchmark dataset named WS-FireNet is constructed, and comprehensive experiments are performed on the WS-FireNet dataset. The experimental results demonstrate that our approach is effective in image-level supervised fire detection.