Data Collection
We use Canon EOS 5D Mark IV to capture the data. The resolution of captured images is set to 6720 × 4480. To capture low/normal-light image pairs, the camera was mounted on a sturdy tripod and controlled remotely via a mobile APP. The camera was not touched between the capture process of normallight and low-light images to avoid vibration. For each pair, we first take the normal-light image. Then the low-light image is captured by changing the shutter, exposure time, and ISO to simulate low-light conditions.
Dataset characteristics
2000 pairs
It contains 2,000 image pairs, which is four times the size of the LOL dataset.
No repeated scenes
Different from the existing real scenes dataset, i.e., LOL, there are no repeated scenes in our PNLI dataset, which is more abundant and superior than LOL.
Object categoriesare rich and common
All images in PNLI are collected from considerably more real scenes, which contain both indoor and outdoor scenes. In addition, the object categories in images are rich and common.
Excellent visual quality and clarity
Excellent visual quality and clarity, which might help in learning pixel-level contextual information.
Rich darkness levels
The darkness levels of low-light images in PNLI are rich, and it can truly restore various situations where the actual image brightness is missing due to insufficient ambient light or human operation mistakes. Therefore, it can effectively verify the stability and robustness of our proposed method.