Two-stage framework reconstructs sharp 4D scenes from blurry handheld videos (techxplore.com)

Researchers have introduced MoBluRF, a two-stage framework that reconstructs sharp 4D scenes and enables novel view synthesis (NVS) from blurry monocular videos — a longstanding challenge for Neural Radiance Fields (NeRFs). Motion blur from handheld phones or drones breaks assumptions of static multi-view input, corrupts camera-pose estimation, and degrades geometry. MoBluRF tackles this by first using Base Ray Initialization (BRI) to roughly reconstruct the dynamic 3D scene and refine the initial “base rays” derived from imprecise camera rays, then applying a Motion Decomposition-based Deblurring (MDD) stage. MDD’s Incremental Latent Sharp-rays Prediction (ILSP) decomposes blur into global camera motion and local object motion and incrementally predicts the latent sharp rays that NeRF needs for high-fidelity rendering. Key technical contributions include the motion decomposition strategy, ILSP, and two novel loss terms: one that separates static and dynamic regions without requiring motion masks, and another that enforces geometric accuracy for moving objects. The result is markedly better deblurring, pose consistency, and geometric precision than prior methods across datasets and blur levels. Published in IEEE TPAMI, MoBluRF promises practical gains for consumer capture (sharper 3D content from phones), robotics and drone perception, and virtual/augmented reality by enabling crisp NeRF reconstructions from everyday shaky footage.