LLVIP: Visible-Infrared Paired Low-light Vision Dataset (No Data Mirror -- Adapter Only)

This card describes a DetectionBench dataset adapter for LLVIP. It does NOT host or redistribute the dataset itself -- LLVIP has an explicit non-commercial license with no redistribution grant. See Getting the Data for the official download links.

Disclaimer

DetectionBench is not an official release of LLVIP and does not host any LLVIP images, annotations, or derived files anywhere -- not on Hugging Face, not in this repository.

LLVIP was created by Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou (Beijing University of Posts and Telecommunications), who retain all rights under an explicit, formal license (the official repository's Term of Use and License.md, not just an informal statement). This repository does not claim ownership of any images or annotations.

What this repository provides instead:

1. A DetectionBench dataset adapter that converts an official LLVIP download into DetectionBench's canonical training layout, once you have obtained the data yourself.
2. A helper command, detectionbench-download-dataset --dataset llvip, that prints the official download links (Google Drive and Baidu Netdisk) and can fetch the Google Drive copy automatically via gdown -- see Getting the Data.
3. This banner, generated locally from an LLVIP copy already converted through the adapter, purely to illustrate the dataset's domain and annotation style.

---

Dataset Description

LLVIP is a registered visible/infrared image-pair dataset for pedestrian detection in low-light conditions: 30,976 image pairs (paired visible + infrared frames, pixel-aligned) captured at night, annotated with pedestrian bounding boxes. Its central finding is that infrared imagery is dramatically more useful than visible-light imagery for detection under these conditions -- the paired visible frames are frequently near-black.

This adapter uses the infrared images only. Building a joint RGB+IR detector is a legitimate, separate research direction that DetectionBench's single-image training/eval pipeline does not attempt; the adapter picks the modality that is independently usable for standard detection.

---

Getting the Data

LLVIP is not mirrored here. Get it directly from the authors:

detectionbench-download-dataset --dataset llvip

Resource	Link
LLVIP dataset (Google Drive)	https://drive.google.com/file/d/1VTlT3Y7e1h-Zsne4zahjx5q0TK2ClMVv/view
LLVIP dataset (Baidu Netdisk, access code 14lc)	https://pan.baidu.com/s/1eQO1Is2NPyd-mgmv1Csbfg

Once downloaded, convert it into DetectionBench's canonical layout:

detectionbench-prepare-coco --dataset llvip --raw-dir <downloaded_dir> --output-dir <coco_out>
detectionbench-convert-coco-to-yolo --input-dir <coco_out> --output-dir <yolo_out>

LLVIP ships with no official validation split; the adapter carves a seeded 15% slice out of train.

Note: the adapter expects infrared/{train,test}/*.jpg + Annotations/*.xml under the raw directory, matching the official README's documented layout -- verified against a real download: 10,221 train / 1,804 valid (15% carve of the official 12,025 train images) / 3,463 test, 29,113 / 5,017 / 8,302 person boxes respectively.

Classes (1)

person

---

Dataset Sources

Original Paper

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision

Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, Wenli Zhou

IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021.

Official Resources

• GitHub Repository: https://github.com/bupt-ai-cz/LLVIP
• Homepage: https://bupt-ai-cz.github.io/LLVIP/

---

License

Non-commercial use only, per the official Term of Use and License.md:

• Free for academic and non-academic entities for non-commercial purposes (research, teaching, publications, personal experimentation).
• Attribution required.
• No commercial use of the dataset or derivative work.
• Faces are blurred in places; using the dataset to identify or invade any individual's privacy is explicitly prohibited.
• The license reserves "all rights not expressly granted" -- it does not grant third-party redistribution of the raw dataset, only conditions for using derivative annotations.

Accordingly:

• No Hugging Face mirror of the data is provided or planned.
• The DetectionBench adapter is provided for local, non-commercial research use against a copy you download yourself.
• If you need broader rights, contact the original authors.

---

Citation

If you use this dataset, please cite:

@inproceedings{jia2021llvip,
  title={LLVIP: A Visible-infrared Paired Dataset for Low-light Vision},
  author={Jia, Xinyu and Zhu, Chuang and Li, Minzhen and Tang, Wenqi and Zhou, Wenli},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
  pages={3496--3504},
  year={2021}
}

---

Acknowledgements

We sincerely thank Xinyu Jia, Chuang Zhu, and their co-authors for creating and publicly releasing this valuable low-light visible-infrared benchmark.