3L-OPTRONICS Dataset
Download

Low-Light
level
dataset.

An aligned, physically realistic,
low-light dataset.

Discover project Downloads information
Model of the dark box: enclosure, rotating stage, camera and integrating sphere
Colour (RGB) capture of the forest scene under night illumination
Colour (RGB) capture of the forest scene at a very low illumination level
Wide-field colour (RGB) capture inside the dark box

01 / The project

Realistic data.
Reliable vision.

Captured under standardized nighttime illumination, our dataset recreates realistic outdoor scenarios across defined levels of darkness.

Each acquisition follows controlled lighting conditions, enabling reliable and reproducible evaluation.

Benchmark

Results

08 neural networks · 05 metrics
Network PSNR SSIM LPIPS NIQE Inference time (ms)
SCI37.680.86730.690529.072.08
EnlightenGAN36.660.86760.080124.0458.14
Zero-DCE37.900.86930.778510.4726.21
KinD41.980.92980.027716.0187.22
URetinex-Net++40.950.91580.031120.87254.01
WaveMamba54.800.98080.001321.711322.63
MambaIR51.850.96140.001121.369199.43
RetinexFormer44.640.96870.021017.78541.14

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

02 / In pictures

The project
in seconds.

Preview of the project presentation video
3L-OPTRONICS DatasetsIntroducing the 3L dataset

03 / As-a-service

Your sensor.
Your night.

The dataset published here relies on our standard night levels. The same calibrated bench can be reconfigured
for your own requirements.

01 / Custom dataset

Built around
your sensor.

Beyond the standard night levels,
we produce acquisitions tailored to your sensor.

  • Specific wavelengths (λ)
  • Visible low light
  • Custom night levels
  • Your optics, your sensor
contact@3l-optronics.com
02 / Sensor characterization

How far does it really see?

We evaluate the ability of a sensor to see at night on the
calibrated bench, and compute the associated metrics,
either on our standard night levels
or on the specifications you define.

  • Standard night levels
  • Your own specifications
  • Measured, not estimated
  • Comparable metrics
contact@3l-optronics.com

04 / Inside the dataset

A structure designed
to be reused.

BlockContentPurposeStatus
01ImagesLow-light paired acquisitionsProvide the raw materialAvailable
02ClassesOne folder per illumination level (n0, n1,…): the folder name is the labelEnable supervised training and evaluationAvailable
03MetadataOne JSON file per folder: image name and stage angle in degrees for each image, plus session time, integration time, start and end anglesMake results comparable and reproducibleAvailable
04PaperScientific article and arXiv linkDocument the method and make the dataset citableAvailable

05 / Examples

A few windows
into the data.

Touch to zoom

N1Illumination level 01
N2Illumination level 02
N3Illumination level 03

Actual captures from the dataset,
taken at different illumination levels.

06 / Resources

Ready toexplore?

The links will go live when the dataset and the article are published.

A specific need?

Scientific paperarXiv · method and benchmarkComing soonAvailable Dataset sampleHugging Face · sample imagesComing soonAvailable