Learning View Priors for Single-view 3D Reconstruction

tackle on single-view reconstruction

• ambiguity in possible shapes
• prior knowledge of category
  • a discriminator on input view and random view
• following the diff. renderer
• 3D representation
  • deform a cube
  • texture and shape both defined on cube
• loss
  • recons. loss: diff in multiview
  • view discrimination loss
  • internal pressure loss: loss points outside