It is difficult to reconstruct dynamic images from undersampled data because motion is ignored, producing wildly inaccurate results. Although neural fields provide a continuous and lightweight representation, previous research mostly relied on implicit smoothness. This study uses the optical flow equation for 2D+time computed tomography to improve neural fields using explicit PDE-based motion regularization.
Source :
For further details, visit: https://hackernoon.com/why-neural-fields-beat-grid-based-methods-for-spatiotemporal-imaging
