---

license: apache-2.0

SICE Paired Low Light Image Enhancement Dataset

Overview

The SICE Paired dataset is derived from the original SICE Dataset Part1 available in the Awesome-Low-Light-Enhancement repository. In SICE, each sample consists of multiple images captured at different exposure levels, accompanied by a corresponding ground truth image stored in a dedicated label folder.

Recent developments in unsupervised and unpaired low-light image enhancement (LLIE) have highlighted the value of leveraging paired low-quality images—often created by randomly selecting two low-light samples—to guide model training and to facilitate the computation of supervised metrics during evaluation. Motivated by these advancements, this dataset is structured to support flexible data loading: users can either select two random low-light images (with seed control for reproducibility) or load all available low-light images per sample.

Dataset Composition

• Training Set: Contains paired samples where each sample includes multiple low-light images (selected based on exposure criteria) and the corresponding ground truth image.

• Test Set: Consists of samples reserved for evaluation. For testing, the under-exposure low-light input is selected as follows:

  • If a sample contains 7 images: the 3rd image (corresponding to -1 EV) is used.
  • If a sample contains 9 images: the 4th image (corresponding to -1 EV) is used.

  Similarly, the over-exposure input is chosen (if needed) as the 5th image in a 7-image sample or the 6th image in a 9-image sample.

During dataset preparation, the following metadata was observed in SICE Dataset Part1:

The testing index in Dataset_part1:

4-23 28 31 33-34 37-39 46-52 55-69 75-79 100-103

Note: Samples containing more than 9 images without a clear indication of under- or overexposure were skipped to maintain consistency.

Folder Structure

The dataset is organized as follows:

SICE_Paired/ 
    ├── train/ 
        │ └── samples/ # Paired samples for training 
            │ ├── sample_1/ │ 
                │ ├── label.jpg # Ground Truth image 
                │ ├── low1.jpg # Low-Light input image
                │ ├── low2.jpg # Low-Light input image
                │ └── ... 
            │ ├── sample_2/ │ 
                │ ├── label.jpg # Ground Truth image 
                │ ├── low1.jpg # Low-Light input image
                │ ├── low2.jpg # Low-Light input image
                │ └── ...  

    └── test/ 
        └── samples/ # Paired samples for testing 
            │ ├── sample_2/ │ 
                │ ├── label.jpg # Ground Truth image 
                │ ├── low1.jpg # Low-Light input image
                │ ├── low2.jpg # Low-Light input image
                │ └── ... 

Data Source and Processing

The dataset is constructed using SICE Dataset Part1 from the Awesome-Low-Light-Enhancement repository. Each original sample contains multiple exposure images along with a corresponding ground truth image in a dedicated label folder. The processing steps include:

• Low-Light Image Selection:
  For each sample, the under-exposed image (i.e., the -1 EV image) is selected as the low-light input. Specifically, if a sample contains 7 images, the 3rd image is used; if 9 images, the 4th image is chosen.

• Pairing with Ground Truth:
  The selected low-light image is paired with its corresponding ground truth image from the label folder.

• Dataset Splitting:
  Samples are split into training, and test sets based on predefined indices, ensuring a standardized evaluation protocol.

Code Samples for Data Loading

• For examples on how to build data loaders and perform preprocessing, see load_data_loaders.py.
• For dataset loading and inspection examples, refer to load_dataset.py.

Related Work

Several recent studies have successfully leveraged the SICE dataset for unsupervised low-light image enhancement tasks:

@inproceedings{fu2023learning,
  title={Learning a Simple Low-Light Image Enhancer From Paired Low-Light Instances},
  author={Fu, Zhenqi and Yang, Yan and Tu, Xiaotong and Huang, Yue and Ding, Xinghao and Ma, Kai-Kuang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={22252--22261},
  year={2023}
}

@InProceedings{Jiang_2024_ECCV,
    author    = {Jiang, Hai and Luo, Ao and Liu, Xiaohong and Han, Songchen and Liu, Shuaicheng},
    title     = {LightenDiffusion: Unsupervised Low-Light Image Enhancement with Latent-Retinex Diffusion Models},
    booktitle = {European Conference on Computer Vision},
    year      = {2024},
    pages     = {}
}

Citation

If you use this dataset in your research, please cite the following paper:

@article{Cai2018deep,
  title={Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images},
  author={Cai, Jianrui and Gu, Shuhang and Zhang, Lei},
  journal={IEEE Transactions on Image Processing},
  volume={27},
  number={4},
  pages={2049-2062},
  year={2018},
  publisher={IEEE}
}

Additionally, if you find the dataset card useful in your workflow, please consider citing the dataset card as follows:

@misc{SICE_paired_dataset_card,
  author       = {Omar Khater},
  title        = {SICE Paired Low Light Image Enhancement Dataset Card},
  howpublished = {\url{https://huggingface.co/datasets/okhater/SICE}},
  year         = {2025},
  note         = {Dataset card curated for efficient access to the SICE Paired dataset}
}