Maria Pilligua, David Serrano-Lozano, Pai Peng, Ramon Baldrich, Michael S. Brown and Javier Vazquez-Corral
Computer Vision Center and Universitat Autònoma de Barcelona
Imaging in low-light environments is challenging due to reduced scene radiance, which leads to elevated sensor noise and reduced color saturation. Most learning-based low-light enhancement methods rely on paired training data captured under a single low-light condition and a well-lit reference. The lack of radiance diversity limits our understanding of how enhancement techniques perform across varying illumination intensities. We introduce the Multi-Illumination Low-Light (MILL) dataset, containing images captured at diverse light intensities under controlled conditions with fixed camera settings and precise illuminance measurements. MILL enables comprehensive evaluation of enhancement algorithms across variable lighting conditions. We benchmark several state-of-the-art methods and reveal significant performance variations across intensity levels. Leveraging the unique multi-illumination structure of our dataset, we propose improvements that enhance robustness across diverse illumination scenarios. Our modifications achieve up to 10 dB PSNR improvement for DSLR and 2 dB for the smartphone on Full HD images.
The MILL dataset can be downloaded from HuggingFace or directly from the following links.
| Split | Camera | Resolution | Download Link |
|---|---|---|---|
| MILLs-DSLR | Nikon D5200 | 600x400 | Link (196MB) |
| MILLs-smartphone | Samsung Galaxy S7 | 600x400 | Link (227MB) |
| MILLf-DSLR | Nikon D5200 | 2012x1340 | Link (8.4GB) |
| MILLf-smartphone | Samsung Galaxy S7 | 1560x1040 | Link (915MB) |
The MILLs splits are the low-resolution splits of the MILL dataset, while the MILLf are the high-resolution ones.
Both MILLs images are resized to 600x400 and follow the following data structure:
MILLs-{split}
|- 📁 train
| |- 📁 input
| | |- 🖼️ Scene1_a-1.png
| | |- 🖼️ Scene1_a-2.png
| | | ...
| | |- 🖼️ Scene1_a-10.png
| | |- 🖼️ Scene1_b-1.png
| | |- 🖼️ Scene2_a-7.png
| | |- 🖼️ Scene9-4.png
| |
| |- 📁 gt
| |- 🖼️ Scene1_a.png
| |- 🖼️ Scene1_b.png
| |- 🖼️ Scene2_a.png
| |- 🖼️ Scene9.png
|
|- 📁 validation
| ...
|- 📁 test
...
The first part of the filename defines the scene name while the second part of the input filenames defines the intensity level (from 1 to 10).
The full resolution DSLR split crops the 6036x4020 original images into 9 2012x1340 images. The filenames have the grid position (row-column id) in between the scene name and intensity level.
MILLf-DSLR
|- 📁 train
| |- 📁 input
| | |- 🖼️ Scene1_a-00-1.png
| | |- 🖼️ Scene1_a-01-1.png
| | | ...
| | |- 🖼️ Scene1_a-22-1.png
| | |- 🖼️ Scene1_a-00-2.png
| | | ...
| | |- 🖼️ Scene1_a-00-10.png
| | |- 🖼️ Scene1_b-22-1.png
| | |- 🖼️ Scene2_a-12-7.png
| | |- 🖼️ Scene9-02-4.png
| |
| |- 📁 gt
| |- 🖼️ Scene1_a-00.png
| |- 🖼️ Scene1_a-01.png
| |- 🖼️ Scene1_a-22.png
| |- 🖼️ Scene1_b-22.png
| |- 🖼️ Scene2_a-12.png
| |- 🖼️ Scene9-02.png
|
|- 📁 validation
| ...
|- 📁 test
...
It follows the same structure as the MILLs splits. The images are the original ones captured by the Samsung Galaxy S7 device.
We benchmark several state-of-the-art low-light enhancement methods on the MILL dataset and reveal significant performance variations across different intensity levels. Our proposed improvements achieve up to 10 dB PSNR improvement for DSLR and 2 dB for smartphone images on Full HD resolution. For detailed results, please refer to our paper.
If you use the MILL dataset or our methods in your research, please cite our paper:
@inproceedings{pilligua2026mill,
title={MILL: Evaluating Low-Light Image Enhancement Across Multiple Intensity Levels},
author={Pilligua, Maria and Serrano-Lozano, David and Peng, Pai and Baldrich, Ramon and Brown, Michael S. and Vazquez-Corral, Javier},
booktitle={CVPR Findings},
year={2026}
}