RE2L: A Real-World Dataset for Outdoor Low-Light Image Enhancement

Low- and normal light image pairs are usually required for training low-light image enhancement (LLIE) deep learning models. Although such training pairs can be synthesized by adjusting exposure, there is currently no outdoor real-world LLIE benchmark. Therefore, it is un-clear how to optimize various LLIE models for real out-door scenes. To fill the gap and benchmark outdoor LLIE tasks, we propose the REal-world two-dimensional Low-light dataset, dubbed RE2L. Natural-light illumination variations over both time and exposure level are fully captured in RE2L, which provides more information on illumination as restoration guidance during model training. We bench-marked the state-of-the-art LLIE methods over the proposed RE2L and its derivation RE2L-P datasets. Furthermore, we showed that using RE2L and RE2L-P for training LLIE models can improve their effectiveness both quantitatively and visually.