ELLAR: An Action Recognition Dataset for Extremely Low-Light Conditions with Dual Gamma Adaptive Modulation

Minse Ha, Wan-Gi Bae, Geunyoung Bae, Jong Taek Lee

Abstract: In this paper, we address the challenging problem of action recognition in extremely low-light environments. Currently, available datasets built under low-light settings are not truly representative of extremely dark conditions because they have a sufficient signal-to-noise ratio, making them visible with simple low-light image enhancement methods. Due to the lack of datasets captured under extremely low-light conditions, we present a new dataset with more than 12K video samples, named Extremely Low-Light condition Action Recognition (ELLAR). This dataset is constructed to reflect the characteristics of extremely low-light conditions where the visibility of videos is corrupted by overwhelming noise and blurs. ELLAR also covers a diverse range of dark settings within the scope of extremely low-light conditions. Furthermore, we propose a simple yet strong baseline method, leveraging a Mixture of Experts in gamma intensity correction, which enables models to be flexible and adaptive to a range of low illuminance levels. Our approach significantly surpasses state-of-the-art results by \(3.39\%\) top-1 accuracy on ELLAR dataset. The dataset and code are available at https://github.com/knu-vis/ELLAR.