Image Translation by Domain-Adversarial Training

Figure 4

Qualitative performance on the map ↔ aerial photograph translation: (a) input; (b) unconditional GAN; (c) unconditional GAN with pixel loss (CycleGAN [20]); (d) GAN with conditioning on the input image; (e) GAN with conditioning on the unaligned target as well as the feature-wise loss (ours); and (f) supervised pix2pix.

(a)

(b)

(c)

(d)

(e)

(f)