Learning to Reconstruct Shapes from Unseen Classes

problem

• NN to approx. mapping from 2d -> 2.5d -> 3d may be overparametried
  • overfitting to training classes
• geometric projections generalize better

pipeline

• RGB -> depth map
• depth map -> partial spherical map
  • spherical map inpainting
  • proj. back to voxel space
• depth map -> 3d voxels (by projection)
• finetune two voxels