Multiple Instance Learning for CNN Based Fire Detection and Localization

Motivated by the state-of-the-art performance achieved by convolutional neural networks (CNN) in visual detection and classification tasks, CNNs have recently been applied to the visual fire detection problem. In this work, we extend the existing CNN based approaches to fire detection in video sequences by incorporating Multiple Instance Learning (MIL). MIL relaxes the requirement of having accurate locations of fire patches in video frames, which are needed for patch level CNN training. Instead, only frame level labels indicating the presence of fire somewhere in a video frame are needed, substantially alleviating the annotation and training efforts. The resulting approach is tested on a new fire dataset obtained by extending some of the previously used fire datasets with video sequences collected from the web. Experimental results show that the proposed method improves fire detection performance upto 2.5%, while providing patch level localization without requiring patch level annotations.