Chung-Ang University Develops MoBluRF for Sharp 4D Reconstructions

Chung-Ang University researchers have developed MoBluRF, a framework that enables the creation of sharp neural radiance fields (NeRF) from blurry videos, announced in a press release. This advancement addresses the challenges of motion blur in videos captured by handheld devices, which previously hindered accurate 3D scene reconstruction.

MoBluRF operates in two stages: Base Ray Initialization (BRI) and Motion Decomposition-based Deblurring (MDD). The BRI stage reconstructs dynamic 3D scenes from blurry videos, refining the initialization of base rays. The MDD stage then uses these base rays to predict latent sharp rays, improving deblurring accuracy by decomposing motion blur into global camera motion and local object motion.

The framework introduces novel loss functions to separate static and dynamic regions without motion masks and enhance geometric accuracy. MoBluRF outperforms existing methods in both quantitative and qualitative measures, offering robust performance against varying degrees of blur. This innovation allows consumer devices like smartphones to produce sharper and more immersive content, marking a significant advancement in the field of NeRFs.