NeRF Differentiable Forward Maps

This is part of my journey of learning NeRF.

2.3. Differentiable Forward Maps

image-20221208175453557
image-20221208175453557

Differentiable rendering

image-20221208181457315
image-20221208181457315

Volume rendering can render fogs. Sphere rendering only render the solid surface, and needs ground truth supervision.? Neural renderer combines the two.

Differentiability of the rendering function itself

  • BRDF Shading? details later.

Differentiation itself

Design a neural network with higher order derivatives constraints and therefore directly use its derivative.

image-20221208182302568
image-20221208182302568

For example the Eikonal equation forces the neural network has a derivative as 1. Adding the eikonal loss then promises the neural network valid.

Generally, this kind of problems are: the solutions are constrained by its partial derivatives.

Special: Identity Operator

\[ \text{Reconstruction} \rightarrow \hat 1()\rightarrow \text{Sensor domain}\\ \text{Reconstruction} == \text{Sensor domain} \]

Q&A:

  • Can we obtain a neural network in just one forward, without optimization?
  • Can we design special forward maps for specific downstream tasks, eg., classification? Absolutely yes. We can design it to represent a compact representation as the sensor domain. The key idea is to get a differentiable function to map your specific recon and sensor domain.