ExDark: 7,363 annotated low-light images across ten illumination conditions
Exclusively Dark (ExDARK) dataset which to the best of our knowledge, is the largest collection of low-light images taken in very low-light environments to twilight (i.e 10 different conditions) to-date with image class and object level annotations.
At a glance
- What is it?
- The Exclusively Dark dataset collects low-light photographs from very dark environments through twilight, with bounding box annotations on twelve object classes, and ships alongside the SPIC low-light enhancement code. It suits a researcher training or evaluating detection in the dark, and the newly clarified terms mean a for-profit user needs written permission rather than just a citation.
- Who is it for?
- ExDark fits an academic team training or evaluating object detection or low-light enhancement on real dark scenes, and it does not fit a commercial product team without first obtaining written permission, because the terms require prior written permission for commercial use of the licensor-controlled content and written confirmation for any project sponsored by a for-profit entity.
- Can I use it commercially?
- Yes. BSD-3-Clause is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 23 days ago.
- What is it written in?
- Mainly MATLAB, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on October 23, 2026, and from our analysis. They are not legal advice.
Editorial analysis
A dataset built for detection in the dark, not just for viewing
The Exclusively Dark dataset was introduced to facilitate new object detection and image enhancement research particularly in the low-light environment, and it is a collection of 7,363 low-light images from very low-light environments to twilight, which the description calls 10 different conditions, with 12 object classes annotated on both image class level and local object bounding boxes.
The bounding boxes are what distinguish it from a photo collection of dark scenes. A dataset of night photographs is easy to assemble and of limited research value. A dataset of night photographs with object level annotations is a training and evaluation set, and the class list is described as similar to PASCAL VOC, which means it is directly comparable with the standard detection benchmarks of that generation.
The paper behind it was accepted for publication in Computer Vision and Image Understanding, with the citation giving the journal, volume 178, pages 30 to 42, and a DOI. The repository also links the paper as an ArXiv document.
The provenance of the images is worth understanding because it determines what the data can be used for. The description says the collection is, to the best of the authors' knowledge, the largest collection of low-light images taken in very low-light environments to twilight to date, and the rights table in the README refers to underlying photographs, movie frames, and images originating from third-party websites or third-party datasets. So the dataset is assembled from real captured images, frames pulled from film, and pictures taken from elsewhere, and those sources are not uniform.
The primary language of the repository is MATLAB, which reflects the SPIC enhancement code rather than the dataset itself, and the licence field is BSD-3-Clause for the code.
Three sets of terms, not one licence
The licensing section is the most consequential part of this repository, and it is recent and unusually careful.
The README states that ExDark contains different categories of content and that different terms apply to different components, with a table that splits the repository into three rows. The original software code and scripts owned by the ExDark authors fall under the applicable software licence stated for that component, with a pointer to the relevant LICENSE file. The ExDark annotations, bounding boxes, class labels, metadata, dataset splits and related project-created documentation fall under DATASET_TERMS.md. And the underlying photographs, movie frames, images originating from third-party websites or third-party datasets, and other third-party content are not licensed under the ExDark Dataset Research Use Terms unless expressly stated, with separate third-party rights possibly applying.
So the BSD-3-Clause badge covers the code, and it does not cover the photographs. That distinction is easy to miss and it is the difference between a permissive licence and a research-use-only grant.
The academic use terms say the ExDark Dataset Materials, as defined in DATASET_TERMS.md, may be used for non-commercial academic research, education, benchmarking and scholarly publication subject to those terms. And there is a clause that catches a lot of corporate work: where a project is carried out by, with, or under the sponsorship of a for-profit entity, written confirmation from the Licensor is required before relying on the academic-use permission.
Commercial use of the licensor-controlled content requires prior written permission, and nothing in the terms grants commercial or other rights in the underlying third-party images. Permission requests go to Dr. Chee Seng Chan.
Why the terms were tightened in October 2026
The README's change log at the top of the repository shows what happened and when, and the sequence is informative.
The dataset was released on May 29, 2018. It was accepted for publication in CVIU on October 31, 2018. Code for low-light image enhancement was released on June 02, 2019, and the dataset link was updated on September 02, 2022. And on September 30, 2026 the README was updated to add DATASET_TERMS.md, RIGHTS_NOTICE.md and THIRD_PARTY_NOTICES.md.
The rights notice section explains the scope of the change, and it does so with the language a lawyer would choose. Permissions are granted only to the extent that the Licensor owns or is authorised to grant the relevant rights, with a pointer to RIGHTS_NOTICE.md. And then the sentence that matters most: these clarified terms apply prospectively from 01 October 2026 and do not purport to revoke or determine the scope of rights validly granted under an earlier version of the repository.
That last clause is the whole point. Eight years of researchers, papers and trained models were built on the previous terms, and the authors have deliberately not tried to claw those back. The change is a forward-looking clarification for future users rather than a retroactive licence change, which is a considerably more defensible position and one that keeps existing academic citations valid.
The copyright statement is scoped the same way. ExDark annotations, metadata and original project-created dataset content are attributed to Chee Seng Chan for the period 2018 to 2026, which covers the period of the project rather than claiming ownership of the photographs themselves.
SPIC, and the enhancement code that came with the dataset
The dataset repository is also the home of the enhancement method that was published alongside it, and that code sits in a SPIC directory with a set of before-and-after animated GIFs referenced from the README.
The README says the source code for their work on low-light image enhancement is now available, and points readers to the SPIC folder in the repository. The release note from June 02, 2019 marks the point at which that code was published, and the linked paper is the CVIU article that introduced the dataset.
This is worth noting for what it is. A dataset paper and a method paper are usually separate artefacts, and having the method implementation in the same repository as the data means a researcher evaluating an enhancement approach can run the reference implementation against the dataset it was tuned on and against images it was not.
The repository layout reflects the split. The Dataset directory holds the images, the Groundtruth directory holds the annotations, the SPIC directory holds the code, and there is a LICENSE file at the top level, which the rights table points to for the code component specifically. There is no build system file, no package manifest and no test suite in the listing, which is consistent with a MATLAB research code release rather than a distributed library.
The images themselves are shown through an Exdark.gif at the top of the README, giving a sense of the range of conditions in a single animation, and the repository description notes the ten different conditions from very low-light through twilight, which is the axis along which a low-light dataset is most usefully sampled.
Where this dataset stops and a different one has to start
The limits of ExDark are the limits of the use case it was built for, and they are visible in the structure rather than hidden.
The first is that it is a detection dataset with a 2018 vocabulary. Twelve object classes, similar to PASCAL VOC, is a small and now dated label set. A modern detector trained on ExDark is trained on a fixed and fairly generic set of things to find, and a project whose target objects are not in those twelve classes cannot use the annotations as they stand.
The second is the provenance problem again, and it is the one that bites hardest in practice. Because the images come partly from third-party websites and third-party datasets, and because the terms explicitly decline to grant rights in them, a project that trains on ExDark and then ships the model inherits an ambiguity. The annotations were created by the authors and are clear, but the underlying photographs are not the authors' to license. The rights notice says permissions are granted only to the extent the licensor owns or is authorised to grant the relevant rights, which is a candid way of saying that part of the data is on borrowed terms.
The third is the age of the data itself. Real low-light capture conditions shift with sensor technology, and a dataset collected in 2018 reflects the cameras of that period. A model trained on it will be evaluated against capture conditions that no longer exist, and its measured performance will not transfer cleanly to current hardware.
The alternative route, for a project that needs the data commercially or needs a current class set, is to assemble a similar corpus from images the project owns or can license directly, and to use ExDark as the benchmark to compare against rather than as the training set. The BSD-3-Clause licensed code and the paper remain usable as a reference implementation and a citation regardless, subject to the code licence rather than the dataset terms.
Editorial conclusion
ExDark fits an academic team training or evaluating object detection or low-light enhancement on real dark scenes, and it does not fit a commercial product team without first obtaining written permission, because the terms require prior written permission for commercial use of the licensor-controlled content and written confirmation for any project sponsored by a for-profit entity. Before building on it, read DATASET_TERMS.md and RIGHTS_NOTICE.md in the repository, since the clarified terms apply prospectively from 01 October 2026 and the underlying photographs come from third parties whose rights the dataset does not grant.
Frequently asked questions
What is in the ExDark dataset?
It is a collection of 7,363 low-light images captured from very low-light environments through twilight, described as 10 different conditions, with 12 object classes annotated at both image class level and with local object bounding boxes, following a PASCAL VOC style annotation scheme.
Can I use the ExDark dataset commercially?
Not without asking first. Commercial use of the ExDark Dataset Materials controlled by the Licensor requires prior written permission, and where a project is carried out by, with, or under the sponsorship of a for-profit entity, written confirmation from the Licensor is required before relying on the academic-use permission. The underlying third-party images are not covered by these terms at all.
What licence applies to the ExDark code and to the images?
Different terms apply to different components. The original software code and scripts owned by the authors follow the software licence in the relevant LICENSE file, the BSD-3-Clause one at the repository root. The annotations, bounding boxes, class labels, metadata, splits and project-created documentation follow DATASET_TERMS.md, and the underlying photographs, movie frames and third-party content are not licensed under the research use terms unless expressly stated.
Does the ExDark repository include the low-light enhancement code?
Yes. The source code for the work on low-light image enhancement is in the SPIC folder of the repository, released in June 2019 two years after the dataset itself, and the same repository holds the CVIU paper describing the dataset and the method.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/cs-chan-exclusively-dark-image-dataset)Was this useful?