Cooperative Neural Radiance Field for Dynamic Scene Deblurring

Abstract

Recent advancements in 3D dynamic scene reconstruction from monocular videos offer novel possibilities for space-time view synthesis. However, video blurriness caused by defocus or motion commonly seen in real-world scenarios can degrade the synthesis quality. Addressing this challenge, we propose an innovative approach that integrates a deformable blurring kernel with ray origin prediction into neural scene flow fields. This integration effectively models the blurriness inherent in the imaging process. Moreover, we introduce the Cooperative Neural Radiance Field (CoopNeRF), which explicitly simulates defocus blur by incorporating bokeh rendering within the physical blurring framework. Our comprehensive experiments are conducted on six various synthesized datasets, which demonstrate that our method significantly outperforms existing techniques and provides superior quantitative results. Our proposed CoopNeRF not only enhances the synthesis quality but also broadens the applicability of 3D reconstruction techniques in real-world scenarios.