An Unconstrained Dataset for Non-Stationary Video Based Fire Detection

Abstract

Challenging ground truth and standardized metrics are a mandatory requirement for the development and evaluation of computer vision algorithms. Despite the significant amount of publications on video based fire detection research it remains difficult to compare different algorithms due to the lack of common evaluation schemes and evaluation datasets. We address both of these issues by presenting a new dataset of fire videos containing frame by frame annotations which may be used for non-stationary fire detection algorithms evaluation. The dataset includes hand-held, robot attached and drone attached footages and aims to boost the development of fully autonomous fire fighter robots. The presented ground truth and metrics may adapt to any state-of-the-art technique and provide a reliable and unbiased solution to compare them.