ExDark: Exclusively Dark Image Dataset (Object Detection)

Unofficial redistribution of the ExDark (Exclusively Dark Image) dataset, reformatted into a standardized YOLO-compatible directory layout.

Disclaimer

This repository is not an official release of the ExDark dataset.

The ExDark dataset was created by the original authors (Yuen Peng Loh and Chee Seng Chan, Universiti Malaya), who retain all copyright and intellectual property rights. This repository does not claim ownership of any images, annotations, or metadata.

This repository exists for two purposes:

1. To reorganize the dataset into a standardized YOLO/Ultralytics-compatible directory structure that can be used directly by many modern object detection training pipelines.
2. To provide a more reliable download source, as the original hosting may be slow, difficult to access, or subject to broken configuration files (see Changes from the Official Release below).

Two-hop provenance. This redistribution is not sourced directly from the original authors' raw distribution. It is sourced from a third-party YOLO-format export of ExDark published on Roboflow Universe, which itself reorganized the original per-class-folder ExDark release into a YOLO-compatible layout. Both the original authors and the intermediate contributor are credited below.

---

Dataset Description

ExDark (Exclusively Dark Image Dataset) is a low-light robustness benchmark: 7,344 images captured across 10 low-light conditions, ranging from very low light to twilight, with both image-level class labels and object-level bounding box annotations across 12 object classes (a subset similar in spirit to PASCAL VOC).

This repository preserves the dataset's images and labels while packaging them in a standardized YOLO directory layout for improved compatibility with modern deep learning frameworks.

---

Changes from the Official Release

The transformation chain has two hops, and each is scoped narrowly:

1. Original ExDark → Roboflow YOLO export (not performed by this repository)

A third-party Roboflow contributor (workspace my-workspace-ohnbt, project exclusively-dark-image, version 2) converted the original ExDark release (per-class image folders + object-level annotation files) into a YOLO-compatible train/valid/test layout, resizing images to 640x640. This step was not performed by us; we redistribute its output.

2. Roboflow YOLO export → this repository

• Fixed a broken data.yaml: the Roboflow-generated file uses a relative path: ../train/images-style reference with inconsistent casing that does not resolve once the directory is placed inside another project. A corrected data.yaml with an explicit, correctly-cased root path is provided in this repository. The original file's train/val/test keys and class list are otherwise unchanged.
• No images were added, removed, or modified.
• No labels were changed.
• No splits were changed.

Apart from the data.yaml path fix, the dataset contents in this repository are identical to the Roboflow export described above, which is itself a direct reorganization of the original ExDark release.

---

Dataset Structure

dataset/
├── README.md
├── data.yaml
├── train/
│   ├── images/
│   └── labels/
├── valid/
│   ├── images/
│   └── labels/
└── test/
    ├── images/
    └── labels/

where:

• images/ contains the RGB low-light images for each split (640x640, pre-resized by the Roboflow export).
• labels/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized).
• data.yaml is the Ultralytics dataset configuration file (class names, split paths).
• Splits: train 5,874 images · valid 736 images · test 734 images (7,344 total).

Classes (12)

Bicycle, Boat, Bottle, Bus, Car, Cat, Chair, Cup, Dog, Motorbike, People, Table

Class distribution is imbalanced: People accounts for roughly 46% of all annotated boxes, while Bus is the rarest class. Keep this in mind when interpreting per-class metrics.

---

Dataset Sources

Original Paper

Getting to Know Low-light Images with The Exclusively Dark Dataset

Yuen Peng Loh, Chee Seng Chan

Computer Vision and Image Understanding, Volume 178, 2019, Pages 30-42.

DOI: https://doi.org/10.1016/j.cviu.2018.10.010

Official Resources

• Official Repository: https://github.com/cs-chan/Exclusively-Dark-Image-Dataset
• Paper (DOI): https://doi.org/10.1016/j.cviu.2018.10.010
• Paper arXiv: https://arxiv.org/abs/1805.11227

Intermediate YOLO Export

• Roboflow Project: my-workspace-ohnbt/exclusively-dark-image (version 2)
• URL: https://universe.roboflow.com/my-workspace-ohnbt/exclusively-dark-image/dataset/2

---

Attribution

All credit for the dataset belongs entirely to the original ExDark authors, Yuen Peng Loh and Chee Seng Chan.

Credit for the YOLO-format reorganization used as the direct source for this repository belongs to the Roboflow contributor at workspace my-workspace-ohnbt.

This repository only redistributes that YOLO-format export, with one configuration-file fix, for improved usability and accessibility.

If you use this dataset in your research, please cite the original publication below.

---

License

The original ExDark dataset is distributed by its authors under the BSD 3-Clause License (see the official LICENSE file), a permissive license requiring attribution and carrying no warranty.

Important — read before commercial use. Separately from the license text itself, the original authors' repository states: "for commercial purpose usage, please contact Dr. Chee Seng Chan" (cs.chan at um.edu.my). This repository honors that request: this redistribution is intended for non-commercial research use, consistent with the authors' stated wishes, even though it is not a term written into the BSD-3-Clause license text itself. If you intend commercial use, contact the original authors directly.

Note on the intermediate Roboflow export's license label. The Roboflow project this repository redistributes from self-declares a "CC BY 4.0" license on its listing page. That label reflects the Roboflow uploader's own claim, not the original copyright holder's terms. This card treats the original authors' official LICENSE file (BSD-3-Clause) and their stated non-commercial request as authoritative, and this repository is distributed under those same terms.

This repository is distributed under the same terms as the original: BSD-3-Clause, with the authors' non-commercial-use request honored.

---

Citation

If you use this dataset, please cite:

@article{Exdark,
  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},
  doi = {https://doi.org/10.1016/j.cviu.2018.10.010}
}

---

Acknowledgements

We sincerely thank Yuen Peng Loh and Chee Seng Chan for creating and publicly releasing this valuable low-light robustness benchmark, and the Roboflow contributor at workspace my-workspace-ohnbt for the original YOLO-format reorganization this repository redistributes.