Here briefly talk about the three papers such as the title of:
Reference: https://blog.csdn.net/gdymind/article/details/82696481
(1) pix2pix: another picture generated from a picture
pipeline follows generator which is U-net;
(2) Cycle the GAN: pix2pix pictures in pairs training, CycleGAN solve this problem, not a right, but similar!
pipeline as follows:
（3）pix2pix HD
Reference: https://www.jianshu.com/p/eb29a264c71a
pipeline as follows,
FIG on the Image generator network G specific structure is as follows:
Total flow: the original low resolution input RGB image, on the one hand to obtain its Labels (semantic labels corresponding semantic tags + Boundary label), after the input on the other hand to give a characteristic diagram obtained Encoder according Labels do average pooling instance- wise feature map, both with and enter into the G generation performed, the generator consists of two parts, G1, and G2, wherein G2 is in turn split into two portions. Pix2pix generator G1 and there is no difference, the structure is a U-Net of end2end. G2 left half of extracted features and features of the previous layer and output layer G1 are summed information fusion, the fusion of information into the second half of the output high-resolution image G2. Determined using multiscale discriminator performs discrimination and the results averaged over three different scales. The three dimensions is determined as follows: the original image, the original image 1/2 down sampling, 1/4 downsampling the original image. Obviously, the coarser scales receptive field, the more concerned about global consistency.
(Labels)