Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling

what makes a good generative model for 3D models?

• varied and specific: fine details
• beyond memorizing and recombining

our model

• distangled representations for generative and discriminative models
• 200D latent vector
• all-conv nn
• discriminator learns faster than generator
  • 3D latent space hard to learn
  • discriminator updated only if acc. <= 80%

“distangle”: by experiment, not by design