Comparison of different methods on the Cityscapes dataset

• pix2pix is trained on aligned pairs.
• “feature loss + GAN” uses the GAN loss plus the L1 difference of the activation of ReLU4_2 of a pretrained VGG network.
• “pixel loss + GAN” uses the GAN loss plus the L1 difference between the pixels of the input and output images.

Input	BiGAN	CoGAN	feature loss + GAN	pixel loss + GAN	CycleGAN	pix2pix	Ground Truth