Visual Object Networks Image Generation with Disentangled 3D Representation

shape code -> voxel grid -> depth map + silhouette map -> 2D image

• first step: vanilla gan
• second step: viewpoint estimation from real data
• last step: cyclegan-ish
  • one-to-many mapping: latent code ignored in one-to-one assumption
  • add one loss: E_text(G_text(v_2.5d, z_text)) ~= z_text
  • and L_KL to let E_text close to a Gaussian