This dataset accompanies the following publication: "RE2L: A Real-World Dataset for Outdoor Low-Light Image Enhancement", IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR), 2022.
Download: OneDrive Link
In this work, we propose a novel REal-world 2-dimentional Low-light (RE2L) dataset with naturally changed illumination environment. It is acquired not only based on manually adjusted exposure level (i.e., the exposure dimension in RE2L), but also collecting images at different times of a day to capture the natural light change (i.e., timing dimension) where complex outdoor lighting conditions are well-captured.
As RE2L provides more comprehensive information of illumination variations, models targeting outdoor LLIE will be better guided in training towards state-of-the-art results.
The download link above contains both the original dataset (RE2L) and the paired version (RE2L-P, natural light vs -0.95 exposure).
The figure below demonstrates examples of some selected pairs from RE2L-P. Upper row: normal images; bottom row: low-light images (5% of normal exposure time).
The table below shows quantitative comparisons on LOL and RE2L dataset. Best highlighted in bold.
| Model / Metric | LOL dataset | RE2L dataset | ||
|---|---|---|---|---|
| KinDoriginal KinDRE2L-P |
20.38 21.42 |
0.83 0.85 |
12.92 12.03 |
0.77 0.71 |
| EnlightenGANoriginal EnlightenGANRE2L-P |
13.72 17.87 |
0.61 0.70 |
10.18 16.55 |
0.59 0.77 |
| Zero-DCEoriginal Zero-DCERE2L-P |
14.86 14.59 |
0.59 0.59 |
10.26 9.61 |
0.63 0.62 |
Some enhancements results are presented in the pictures below:
All data is subject to copyright and may only be used for non-commercial research.
Kindly drop 
an email for any questions.
If you find this dataset useful in your research, please cite:
@inproceedings{lin2022,
author={Lin, Yuxin and Li, Jian and Guo, Lanqing and Wen, Bihan},
title={RE2L: A Real-World Dataset for Outdoor Low-Light Image Enhancement},
year={2022},
booktitle={IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR)}
}