Text2Face dataset processing and expansion

[DarkGeekMS]

Process and expand the current Text2Face dataset through the following ideas :

• Main idea :

  • Form pseudo-text descriptions from the CelebA attributes list : @MouhamedAhmed
  • Thus, the text descriptions with be matched with CelebA-HQ faces. Faces re-projection might be required to get unextended latent vectors (W) : @DarkGeekMS @MohamedRamzy1

• Alternative idea :

  • Generate 200,000 faces from StyleGAN2 with their corresponding latent vector.
  • Get the top matches of the 4000 captioned faces from the generated database using FaceNet.
  • Filter the top matches manually :
  • Expand the dataset through :
    • Generate paraphrases for the face captions and use the multiple captions for the same face.
    • Get the closest k matches to a face, in order to have multiple data points from a single caption.