abbaab/DGFF-dataset

DGFF Dataset (LOL + ExDark)

Data used in "End-to-End Task-Oriented Low-Light Image Enhancement via Detection-Guided Feature Feedback" (ICCA 2026), the paper introducing DGFF (Detection-Guided Feature Feedback).

• Code: https://github.com/ShuvroSankar/DGFF
• Companion repository: https://huggingface.co/datasets/abbaab/DGFF

What is in this dataset

DGFF is trained in stages that need two different datasets:

Source	Used for	Size	Split
LOL (Wei et al., 2018)	Phase 1: enhancement training with paired supervision	500 low/normal-light pairs	485 train / 15 validation
ExDark (Loh & Chan, 2019)	Phase 2: adapter fine-tuning; Phase 3: detection training and evaluation	7,363 real low-light images, 12 object classes	2,999 train / 1,800 val / 2,563 test (official imageclasslist.txt splits)

ExDark classes: Bicycle, Boat, Bottle, Bus, Car, Cat, Chair, Cup, Dog, Motorbike, People, Table.

Total repository size is about 5.7 GB. The splits are train, validation and test.

Structure

DGFF-dataset/
├── train/         # image folders
├── validation/    # text-based split lists
└── test/          # image folders

Usage

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="abbaab/DGFF-dataset",
    repo_type="dataset",
    local_dir="./data",
)

or from the command line:

pip install -U huggingface_hub
huggingface-cli download abbaab/DGFF-dataset --repo-type dataset --local-dir ./data

Then follow the training and evaluation instructions in the GitHub repository.

Evaluation protocol

The paper uses a domain-matched protocol on ExDark: separate YOLOv8n detectors are trained on raw, LLEN-enhanced and DGFF-enhanced training images (50 epochs each) and each is evaluated on test images from its own distribution. This avoids the domain mismatch that disadvantages enhanced images in conventional evaluation.

Pipeline	mAP@0.5	mAP@0.5:0.95
Raw (baseline)	0.408	0.232
LLEN only	0.469	0.276
DGFF	0.474	0.280

Licensing and attribution

This repository redistributes images from existing public datasets. The original licences and terms of use of LOL and ExDark apply, and are typically restricted to non-commercial academic research. Please check the original sources before any other use and cite them.

• LOL: https://daooshee.github.io/BMVC2018website/
• ExDark: https://github.com/cs-chan/Exclusively-Dark-Image-Dataset

Citation

@inproceedings{sen2026dgff,
  title     = {End-to-End Task-Oriented Low-Light Image Enhancement via Detection-Guided Feature Feedback},
  author    = {Sen, Shuvro Sankar and Mia, MD. Maruf and Hasan, Naim and Chayon, Muhammad Hasibur Rashid},
  booktitle = {Proceedings of The 4th International Conference on Computing Advancements (ICCA 2026)},
  year      = {2026},
  address   = {Dhaka, Bangladesh},
  publisher = {ACM}
}

@inproceedings{wei2018deep,
  title     = {Deep Retinex Decomposition for Low-Light Enhancement},
  author    = {Wei, Chen and Wang, Wenjing and Yang, Wenhan and Liu, Jiaying},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2018}
}

@article{loh2019getting,
  title   = {Getting to Know Low-Light Images with the Exclusively Dark Dataset},
  author  = {Loh, Yuen Peng and Chan, Chee Seng},
  journal = {Computer Vision and Image Understanding},
  volume  = {178},
  pages   = {30--42},
  year    = {2019}
}

Contact

Shuvro Sankar Sen, American International University - Bangladesh (AIUB). Issues and questions: use the Community tab or the GitHub repository.