The challenge on
Low-light Raw Video Denoising
With Realistic Motion

Data Collection

As shown in Algorithm 1, we first collect 70 high-quality 4k videos from the internet, then play them on the DELL U2720QM monitor. We use a Sony Alpha 7R IV full-frame mirrorless camera. The size of the Bayer image is 9504×6336. The scenes of the video clips contain indoor and outdoor, ranging from natural landscapes to extreme sports. This relatively large range of scenes also has an advantage compared to previous datasets. Examples of our data are in Figure 1.

Dataset characteristics

Realistic scene motion
We collect paired low-light raw videos with realistic motion, showing great generalization to the complex scenarios in the real world.

210 clips
It contains 210 video pairs, each scene contains three noise levels.

High-quality ground truth
Previous datasets are all collected in degraded conditions, which may significantly decline the performance of the network trained on them when tackling real scenes. We directly obtain realistic motion in our raw low-light video denoising dataset, featuring high-quality data.

No extra equipment
Our dataset collecting pipeline requires no extra equipment used in previous datasets.