On the generalization of learning-based 3D reconstruction

learn a scalar field

• xyz -> occupancy, conditioned on input view
• image -> feature -> alignment loss
  • camera pose given
• um, since camera poses are given, there must exist a more elegant way of incorporating information across different views, like, even geometry based ones would work, instead of relying heavily on the representation capacity of neural nets