Differentiable Volumetric Rendering Learning Implicit 3D Representations without 3D Supervision

kind of like what I have tried before

• point -> occupancy -> surface -> depth map -> surface points -> color
  • problem: how gradient flow through surface? refer to paper for details
• implicit function conditioned on z from image
  • how can that work with NeRF?
    • if given a latent variable extracted (and averaged) from some images, NeRF should be able to extend to generalize on various scenes, which means: no parameter should be scene dependent including params. in CNN and MLP
    • if we do not generalize NeRF, then the additional model capacity from CNN need to be subtracted (somehow), if we are gonna compare that result with NeRF