STANet: a Spatial-Temporal Aggregation Network for Video Deraining

Rain can significantly reduce the visibility of a scene, which hinders a lot of computer vision systems, such as outdoor surveillance and autonomous driving. Comparing with single images, rainy videos can record abundant spatial-temporal information of target scenes, which have tremendous advantages in a wide of computer vision tasks. However, how to effectively extract and aggregate spatial representations and temporal correlations of a video sequence for rain removal is a nontrivial task. In this paper, we propose a novel spatial-temporal aggregation network, called STANet, for video deraining. It can learn the temporal correlations and spatial representations of consecutive rainy frames, and then adaptively aggregate them through an iterative attention mechanism. Finally, a refinement sub-network is proposed to further remove the remaining rain streaks and generate clean rain-free video. Extensive experiments demonstrate that our proposed network outperforms state-of-the-art methods by delivering excellent results in both quality and speed.