PiLLVD: Real-World Low-Light Video Dataset for Dynamic Scenes

PiLLVD (Pi-based Low-Light Video Dataset) is a paired low-light and reference video dataset designed for research on low-light video enhancement in dynamic scenes. The dataset contains 24 indoor and outdoor sequences comprising 4,531 low-light frames and 4,531 corresponding reference frames. The paired sequences were acquired simultaneously using a synchronised dual-camera acquisition system based on Raspberry Pi devices, Sony IMX519 camera modules, and a 70/30 beam splitter. The acquisition setup enables low-light and corresponding reference observations of the same dynamic scene to be captured under different illumination conditions. The dataset includes scenes containing motion from people, objects, or displayed content while the cameras remain stationary. Post-capture spatial alignment was applied to improve correspondence between the paired low-light and reference frames. PiLLVD is intended to support the development, evaluation, and benchmarking of low-light video enhancement methods using paired real-world video data.