$\mathrm 3D^2Unet$: 3D Deformable Unet for Low-Light Video Enhancement.

Video recording suffers from noise, artifacts, low illumination, and weak contrast under low-light conditions. With such difficulties, it is challenging to recover a high-quality video from the corresponding low-light one. Previous works have proven that convolutional neural networks perform well on low-light image tasks, and these methods are further extended to the video processing field. However, existing video recovery methods fail to fully exploit the long-range spatial and temporal dependency simultaneously. In this paper, we propose a 3D deformable network based on Unet-like architecture (\(\mathrm 3D^2Unet\)) for low-light video enhancement, which recovers RGB formatted videos from RAW sensor data. Specifically, we adopt a spatial temporal adaptive block with 3D deformable convolutions to better adapt the varying features of videos along spatio-temporal dimensions. In addition, a global residual projection is employed to further boost learning efficiency. Experimental results demonstrate that our method outperforms state-of-the-art low-light video enhancement works.