View Independent Generative Adversarial Network for Novel View Synthesis

encoder-decoder architecture

• encoder: view-independent feature
• decoder: feature with user specific viewpoints
• …and more
  • discriminators: classify real/gen images; pose estimation of input image
• ability to gen views outside the training set
  • 3d structure in latent space
• input: img[:3] + camera_pose[:12], camera pose provided for each pixel…
• coord conv instead of standard conv: concate pixel location
• and adain regularizer… fancy
• and yet tons of regularizer…gg
  • how are these back propagated? need src code
• encoder help 3d learning
  • feature from encoder feed into other 3d tasks