ResFields

ResFields are a class of networks specifically designed to effectively represent complex spatiotemporal signals. It introduces time-varying weights into multi-layer perceptrons and uses trainable residual parameters to enhance the expressive ability of the model. The method can be seamlessly integrated into existing technologies and can significantly improve results on a variety of challenging tasks, such as 2D video approximation, dynamic shape modeling, and dynamic NeRF reconstruction.

Demand group:

["2D Video Approximation", "Dynamic Shape Modeling", "Dynamic NeRF Reconstruction"]

Example of usage scenario:

for video compression and reconstruction

For modeling and rendering of dynamic 3D scenes

For the capture and reconstruction of time-varying 3D data

Product features:

Introducing time-varying weights into multilayer perceptrons

Utilize trainable low-rank residual parameters to enhance model expressiveness

Seamlessly compatible with existing MLP networks, maintaining inference and training speeds

Improve the generalization ability of the model

Can be widely used in various spatiotemporal signals represented by MLP networks