MoBluRF: Sharp 4D neural radiance field reconstructions from blurry monocular videos captured from everyday handheld devices

https://www.eurekalert.org/news-releases/1098837

"3D from differently angled 2D images... neural network predicts color/ density at any point in 3D space, casting imaginary light rays from camera through each pixel in all input images, sampling points along rays with 3D coordinates/ viewing direction... uses motion decomposition, avoiding mask supervision... Base Ray Initialization roughly reconstructs dynamic 3D scenes/ Motion Decomposition-based Deblurring predicts latent sharp rays incrementally decomposes into global motion/ local object motion components"