Answers are grounded in this paper and cite their sources.
Reconstructing a sharp dynamic radiance field from blurry monocular video is an important problem in novel view synthesis (NVS). Recent approaches jointly optimize dynamic scene representations with blur formation models. However, these methods often suffer from unstable optimization and struggle to recover sharp results under severe blur. To alleviate this problem, we propose a sequential pipeline that combines neural network-based video deblurring with dynamic radiance field optimization. Our method first applies initial deblurring to the input video and learns a dynamic radiance field from the deblurred frames. It then renders images from the learned radiance field and uses them as guidance for a RF-guided video deblurring model to restore the input video again. Finally, the restored sharp frames are used to optimize the final dynamic radiance field. Experimental results show that the proposed method achieves higher PSNR and lower LPIPS than existing methods in novel view synthesis, while demonstrating robust restoration performance in regions with severe motion blur.
No takes yet. Share an insight, caveat, or question.
권강민 et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: