Low-light is More Than Darkness: An Empirical Study on Illumination Types and Enhancement Methods

Low-light images challenge both human perception and computer vision algorithms. Despite notable progress in this field, there are still various gaps that are yet to be investigated, such as the significance of low-light illumination characteristics towards image enhancement and object classification. Therefore, this paper details various analyses to study this phenomenon and provide insights for future developments of algorithms and solutions. Specifically, comparative analysis was done to investigate human and machine perception towards "low-light types", followed by empirical studies on the effect of illumination types towards state-of-the-art image enhancement quality and also their pre-processing capability for downstream task, namely object classification. It is found that illumination types significantly influences the performance of enhancement algorithms that tend to cater for a "general" type of low-light illumination. This lack of illumination type awareness therefore leads models to perform well in certain conditions, but severely underperforms in others. Thus, it is imperative for upcoming works to incorporate such illumination information for potential breakthroughs in this area.