This document presents a new unconstrained dataset for evaluating non-stationary video-based fire detection algorithms. The dataset contains 24 videos totaling over 28,000 frames annotated for fire regions. This is the first publicly available dataset of its kind, containing a variety of fire sources, camera movements, and other challenges. The document outlines metrics for evaluating fire detection algorithms frame-by-frame and based on location. It also provides sample results comparing previous algorithms on the new dataset. The goal is to enable independent evaluation of fire detection techniques and progress in this area of research.
Related topics: