Photo-geometric autoencoding to learn 3D objects from unlabelled images

learn from unlabelled data

• minimal assumptions
• previous works: with priors, monocular learning is possible
  • what prior?
• learn from unconstrained images
  • diff. views of diff. identities
• insights
  • var. in images gen.ed by vp. geometry and texture
  • registeration of vp. reduces var. significantly
• network
  • image -> vp / depth map / light / albedo -> rendered image –> loss
• gonna re-read later…confused