Few-Shot Generalization for Single-Image 3D Reconstruction via Priors

category priors

• prior shape + input image
• sparsity of output space -> lack of data
  • need priors!
• generalize to novel object with no additional training
  • generatizability of NN rather than fewshot method
• so… quite similar to conditional U-Net
  • with iterative refinement
  • in multiview case: one new image in one step