Variable Grey asks how well do generative adversarial networks (GANs) learn deeper topology structures that match images?

The concept of figure-ground originated in the 14th century. It has since undergone gradual improvement and standardisation. Its theoretical scope has been continuously expanded and reinterpreted. The simple black and white space division gradually became ambiguous and abundant in layers; however few studies have further analysed the grey spatial hierarchy and the specific functions and interfaces corresponding to each level.

GANs are relatively developed at learning 2D images, but the problem of making computers understand topology has been one to overcome in the field of machine learning. This project focuses on experiments and discussions on the effect of a specific GAN called pix2pix in learning the deep topology of images. This problem is analysed in two stages: First, how well does pix2pix image-to-image translation, learn the topology of the image? Second, what is the effect of greyscale and RGB mode on the process of pix2pix learning topology?