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

less than 1 minute read

Published:

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