Driving-Video Dehazing with Non-Aligned Regularization for Safety Assistance

Real driving-video dehazing poses a significant challenge due to the inherentdifficulty in acquiring precisely aligned hazy/clear video pairs for effectivemodel training, especially in dynamic driving scenarios with unpredictableweather conditions. In this paper, we propose a pioneering approach thataddresses this challenge through a nonaligned regularization strategy. Our coreconcept involves identifying clear frames that closely match hazy frames,serving as references to supervise a video dehazing network. Our approachcomprises two key components: reference matching and video dehazing. Firstly,we introduce a non-aligned reference frame matching module, leveraging anadaptive sliding window to match high-quality reference frames from clearvideos. Video dehazing incorporates flow-guided cosine attention sampler anddeformable cosine attention fusion modules to enhance spatial multiframealignment and fuse their improved information. To validate our approach, wecollect a GoProHazy dataset captured effortlessly with GoPro cameras in diverserural and urban road environments. Extensive experiments demonstrate thesuperiority of the proposed method over current state-of-the-art methods in thechallenging task of real driving-video dehazing. Project page.