By requesting access, you agree to the 3L-Dataset Research License: non-commercial research use only, no redistribution of the images (original or modified), and citation of the paper in any publication.

3L-Dataset · Low Light Level

An aligned, physically realistic, low-light dataset.

Developed by 3L-OPTRONICS, the 3L-Dataset provides paired captures of the same scenes under standardized nighttime illumination levels, for training and evaluating Low-Light Image Enhancement (LLIE) methods.


🌐 Website	https://3l-d.github.io/3L-Dataset/
📄 Paper	3L-D: A Low-Light Level Dataset from a Sensor-Agnostic, Alignment-Preserving Acquisition Framework (pending, arXiv link coming soon)
✉️ Contact	contact@3l-optronics.com
🤗 Dataset	https://huggingface.co/datasets/3L-OPTRONICS/3L-Dataset
💼 LinkedIn	https://www.linkedin.com/company/3l-optronics
⚖️ License	3L-Dataset Research License: non-commercial research only

This repository hosts a sample of the dataset. See the paper, 3L-D: A Low-Light Level Dataset from a Sensor-Agnostic, Alignment-Preserving Acquisition Framework, for the full acquisition protocol.

⚠️ License at a glance

This dataset is released under the 3L-Dataset Research License. By downloading it, you agree to its terms:

✅ Non-commercial research and teaching only
❌ No commercial use of the dataset or of models trained on it
❌ No redistribution of the images, in original or modified form
📝 Cite the paper in any publication using the dataset

Full terms in License and LICENSE.md. Commercial use: contact@3l-optronics.com

Why this dataset

Most low-light datasets rely on synthetic darkening or on handheld captures that are never perfectly aligned. The - is captured on a calibrated optical bench:

Physically realistic: darkness is produced by real, measured illumination, not simulated in post-processing.
Pixel-perfect alignment: camera and scene stay fixed between levels, so every low-light image has an exactly aligned reference.
Reproducible: each acquisition follows the same controlled lighting conditions.

Because full-reference metrics such as PSNR, SSIM and LPIPS compare images pixel by pixel, this alignment makes their scores interpretable: they measure the quality of the restoration, not a camera shift.

Acquisition setup

Light-tight enclosure with an integrating sphere for uniform, calibrated illumination
Rotating stage to capture each scene from several angles
Scale-model scenes reproducing outdoor environments (countryside, forest, desert, urban)
Fixed camera and optics (f/1.2 objective)
6 illumination levels per scene and per position: one reference level (gt) and five night levels (n1 to n5)

Dataset structure

This sample contains the reference level and three night levels:

3l-dataset/
├── metadata.jsonl      ← one line per image, read by load_dataset
├── gt/                 ← reference (ground truth)
│   ├── metadata.json
│   ├── 0001.png
│   └── ...
├── n2/
│   ├── metadata.json
│   └── ...
├── n3/
└── n4/

Folder	Content
`gt`	Ground truth: reference capture of each scene and position, used as the target for full-reference metrics
`n2`, `n3`, `n4`	The same scenes and positions under increasingly low night illumination

gt is the reference level (named n0 on the acquisition bench). Night levels n1 and n5 are part of the full dataset but not of this sample. For a given scene and position, the captures of all levels are aligned pixel by pixel.

Block	Content	Purpose
Images	Low-light paired acquisitions	Provide the raw material
Classes	One folder per illumination level (`gt`, `n2`, `n3`, `n4`): the folder name is the label	Enable supervised training and evaluation
Metadata	One JSON file per folder: image name and stage angle (degrees) for each image, plus session time, integration time, start and end angles. The same information is gathered in `metadata.jsonl` at the root	Make results comparable and reproducible

Image format: 16-bit grayscale PNG, 2808 × 1096 px.

16-bit depth preserves the full dynamic range of the sensor, which matters in very low light where most of the signal sits in the lowest grey levels. Load the images without converting them to 8-bit:

import numpy as np
from PIL import Image

img = np.array(Image.open("gt/0001.png"))   # dtype uint16, shape (1096, 2808)
img = img.astype(np.float32) / 65535.0       # normalized to [0, 1]

Metadata example

Excerpt of gt/metadata.json:

{
  "session_time": "2026-08-21T10:00",
  "integration_time_ms": 120,
  "start_angle_deg": 0,
  "end_angle_deg": 360,
  "images": [
    { "image": "gt/0001.png", "angle_deg": 0 },
    { "image": "gt/0011.png", "angle_deg": 150 }
  ]
}

Usage

from datasets import load_dataset

ds = load_dataset("3L-OPTRONICS/3L-Dataset")
print(ds)

This dataset is gated: request access on this page first, then log in once from your terminal before loading it:

hf auth login

load_dataset reads metadata.jsonl automatically: every image comes with its illumination level (level), its stage angle and its session information.

sample = ds["train"][0]
sample["image"]        # PIL image
sample["level"]        # "gt", "n2", "n3" or "n4"
sample["angle_deg"]    # stage angle in degrees

The per-folder metadata.json files remain available for direct reading:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("3L-OPTRONICS/3L-Dataset", "gt/metadata.json", repo_type="dataset")
meta = json.load(open(path))

Benchmark

Eight LLIE networks evaluated on the 3L-Dataset (see the paper for the full protocol).

Network	PSNR ↑	SSIM ↑	LPIPS ↓	NIQE ↓	Inference time (ms) ↓
SCI	37.68	0.8673	0.6905	29.07	2.08
EnlightenGAN	36.66	0.8676	0.0801	24.04	58.14
Zero-DCE	37.90	0.8693	0.7785	10.47	26.21
KinD	41.98	0.9298	0.0277	16.01	87.22
URetinex-Net++	40.95	0.9158	0.0311	20.87	254.01
WaveMamba	54.80	0.9808	0.0013	21.71	1322.63
MambaIR	51.85	0.9614	0.0011	21.36	9199.43
RetinexFormer	44.64	0.9687	0.0210	17.78	541.14

PSNR and SSIM are higher-is-better; LPIPS, NIQE and inference time are lower-is-better.

Custom datasets and sensor characterization

The same calibrated bench can be reconfigured for your own requirements:

Custom datasets: specific wavelengths (λ), visible low light, custom night levels, your optics and sensor.
Sensor characterization: evaluate how well a sensor sees at night, with metrics computed on our standard night levels or on your own specifications.

Contact: contact@3l-optronics.com

License

The 3L-Dataset is released under the 3L-Dataset Research License.


✅ Allowed	Download and use the images for non-commercial research and teaching
✅ Allowed	Train and evaluate models for non-commercial research
✅ Allowed	Show a limited number of images and processed results (e.g. enhanced outputs) in scientific publications, with citation
❌ Prohibited	Any commercial use: products, services, paid consulting, training or benchmarking commercial models
❌ Prohibited	Redistributing the dataset or any part of it, in original or modified form (crops, edits, re-encoding, mirrors, other dataset hubs)
❌ Prohibited	Creating and sharing a derived dataset from these images
❌ Prohibited	Removing copyright notices or misrepresenting the origin of the images
📝 Required	Citing the 3L-Dataset paper in any publication or presentation using the dataset

Access to this repository is gated: you must accept the license and provide your affiliation before downloading.

Commercial use, redistribution or custom datasets are available under a separate agreement: contact@3l-optronics.com

Citation

If you use the 3L-Dataset, please cite:

@misc{rueff2026_3ld,
  title  = {{3L-D}: A Low-Light Level Dataset from a Sensor-Agnostic, Alignment-Preserving Acquisition Framework},
  author = {Rueff, L{\'e}o and Ghezal, Maurg{\`e}ne and Gorse, Valentin and Dufaure, Astrid and
            Monsalvo, Esteban and Atteia, Fran{\c{c}}ois and Escoubas, Ludovic},
  year   = {2026},
  note   = {Preprint, arXiv submission in preparation}
}

Downloads last month: -

Total file size:

1.74 GB