Skip to content

Latest commit

 

History

History
51 lines (46 loc) · 3.11 KB

Unsupervised_Image-to-Image_Translation_Networks.md

File metadata and controls

51 lines (46 loc) · 3.11 KB

Paper

  • Title: Unsupervised Image-to-Image Translation Networks
  • Authors: Ming-Yu Liu, Thomas Breuel, Jan Kautz
  • Link: https://arxiv.org/abs/1703.00848
  • Tags: Neural Network, GAN
  • Year: 2017

Summary

  • What

    • They present a method to learn mapping functions that transform images from one style to another style. (E.g. photos from daylight to nighttime.)
    • Their method only requires example images for both styles (i.e. class labels per image).
  • How

    • Architecture
      • Their method is based on VAEs (i.e. autoencoders) and GANs.
      • Their architecture is kinda similar to an autoencoder.
      • For an image style A, an encoder E first transform an image to a vector representation z. Then a generator G transforms z into an image.
      • There are two encoders (E1, E2), one per image style (A, B).
      • There are two generators (G1, G2), one per image style (A, B).
      • There are two discriminators (D1, D2), one per generator (and therefore style).
      • An image can be changed in style from A to B using e.g. G2(E1(I_A)).
      • The weights of the encoders are mostly tied/shared. Only the last layers are not-shared.
      • The weights of the generators are mostly tied/shared. Only the last layers are not-shared.
      • They use 3 convs + 4 residual blocks for the encoders and 4 residual blocks + 3 transposes convs for the generators. They use normal convs for the discriminators. Nonlinearities are LeakyReLUs.
      • The encoders are VAEs and trained with common VAE-losses (i.e. lower bound optimization). However, they only predict mean values per component in z, not variances. The variances are all 1.
      • Visualization of the architecture:
        • architecture
    • Loss
      • Their loss consists of three components:
        • VAE-loss: Reconstruction loss (absolute distance) and KL term on z (to keep it close to the standard normal distribution). Most weight is put on the reconstruction loss.
        • GAN-loss: Standard as in other GANs, i.e. cross-entropy.
        • Cycle-Consistency-loss: For an image I_A, it is expected to look the same after switching back and forth between image styles, i.e. I_A = G1(E2( G2(E1(I_A)) )) (switch from style A to B, then from B to A). The cycle consistency loss uses a reconstruction loss and two KL-terms (one for the first E(.) and one for the second).
  • Results

    • When testing on the (satellite) map dataset:
      • Weight sharing between encoders and between generators improved accuracy.
      • The cycle consistency loss improved accuracy.
      • Using 4-6 layers (as opposed to just 3) in the discriminator improved accuracy.
    • Translations that added details (e.g. night to day) were harder for the model.
    • After training, the features from each discriminator seem to be quite good for the respective dataset (i.e. unsupervised learned features).
    • Example translations:
      • examples