16.03.2021 Views

Advanced Deep Learning with Keras

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Cross-Domain GANs

In computer vision, computer graphics, and image processing a number of tasks

involve translating an image from one form to another. As an example, colorization

of grayscale images, converting satellite images to maps, changing the artwork style

of one artist to another, making night-time images into daytime, and summer photos

to winter, are just a few examples. These tasks are referred to as cross-domain

transfer and will be the focus of this chapter. An image in the source domain

is transferred to a target domain resulting in a new translated image.

A cross-domain transfer has a number of practical applications in the real world.

As an example, in autonomous driving research, collecting road scene driving data

is both time-consuming and expensive. In order to cover as many scene variations

as possible in that example, the roads would be traversed during different weather

conditions, seasons, and times giving us a large and varied amount of data. With the

use of a cross-domain transfer, it's possible to generate new synthetic scenes that look

real by translating existing images. For example, we may just need to collect road

scenes in the summer from one area and gather road scenes in the winter from

another place. Then, we can transform the summer images to winter and the winter

images to summer. In this case, it reduces the number of tasks having to be done

by half.

Generation of realistic synthesized images is an area that GANs excel at. Therefore,

cross-domain translation is one of the applications of GANs. In this chapter, we're

going to focus on a popular cross-domain GAN algorithm called CycleGAN [2].

Unlike other cross-domain transfer algorithms, such as a pix2pix [3], CycleGAN

doesn't require aligned training images to work. In aligned images, the training data

should be a pair of images made up of the source image and its corresponding target

image. For example, a satellite image and the corresponding map derived from this

image. CycleGAN only requires the satellite data images and maps. The maps may

be from another satellite data and are not necessarily previously generated from the

training data.

[ 203 ]

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!