Unsupervised shape transformer for image translation and cross-domain retrieval

challenge: input and output do not share same geometrical information

• unpaired data
• CycleGAN: content in input and output images are aligned
• keeping appearance information while transfer shape information
  • object-focused: object in context
  • content space and style space shared
  • above constraints not enough when output is multi-model -> weak adverserial learning
  • attention: literially mask on middle layer
  • AdaIn by style
  • fit-in module: copy-paste in latent space
• losses
  • symmetry loss: inherent prior of clothing and faces