MoBluRF: Motion Deblurring Neural Radiance Fields for Blurry Monocular Video

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural Radiance Fields (NeRF) excel at novel view synthesis for static scenes.
• Motion blur in videos, caused by movement during exposure, poses a significant challenge for synthesizing sharp spatio-temporal views.
• Existing NeRF-based video synthesis methods struggle with motion blur, limiting their effectiveness.

Purpose of the Study:

• To propose a novel motion deblurring NeRF framework, MoBluRF, specifically designed for blurry monocular videos.
• To address the challenge of synthesizing sharp spatio-temporal views from videos affected by motion blur.
• To develop a method that can effectively decompose and handle both global camera motion and local object motion.

Main Methods:

• Introduced MoBluRF, a framework with two stages: Base Ray Initialization (BRI) and Motion Decomposition-based Deblurring (MDD).
• The BRI stage coarsely reconstructs dynamic 3D scenes and initializes base rays for predicting latent sharp rays using inaccurate camera poses.
• The MDD stage employs Incremental Latent Sharp-rays Prediction (ILSP) to decompose latent sharp rays into global and local motion components, utilizing novel loss functions for geometry regularization and scene decomposition without masks.

Main Results:

• MoBluRF successfully synthesizes sharp spatio-temporal views from blurry monocular videos.
• The proposed method effectively decomposes latent sharp rays into global camera motion and local object motion.
• Experiments demonstrate that MoBluRF significantly outperforms state-of-the-art methods both qualitatively and quantitatively.

Conclusions:

• MoBluRF presents a robust solution for motion deblurring in monocular videos using a NeRF-based approach.
• The framework's ability to decompose motion and regularize geometry enables high-quality view synthesis.
• MoBluRF advances the field of video novel view synthesis by effectively handling motion blur.