The content discusses a novel approach to recovering a sharp video sequence from a single motion-blurred image. The key insights are:
Motion blur decomposition is highly ambiguous, as the averaging effects of motion blur have severely destroyed the temporal ordering of latent frames. Existing solutions that rely on neighboring blurry frames or human annotations struggle to resolve this ambiguity.
The authors propose a dual imaging setting that combines a global shutter (GS) blurred image and a rolling shutter (RS) image. The RS view implicitly encodes the temporal ordering of latent frames, which can be leveraged to robustify the motion decomposition.
The authors develop a triaxial imaging system to capture aligned GS blur, RS, and high-speed ground truth video, and construct a real-world dataset called RealBR.
The authors introduce a novel neural network architecture with dual streams for motion interpretation. The blur stream focuses on contextual characterization while the RS stream handles temporal abstraction. A shutter alignment and aggregation module promotes mutual information exchange between the two streams.
Extensive experiments on the RealBR dataset demonstrate the effectiveness of the proposed dual Blur-RS setting and the neural network model in recovering sharp video sequences from single blurred inputs, outperforming state-of-the-art methods.
他の言語に翻訳
原文コンテンツから
arxiv.org
深掘り質問