Evaluating clustering methods underpinning content generation in games using GANs

In recent years there has been a push for more customisation options in games, along with a desire for greater realism. While graphics have been steadily improving year over year, the current customisation options remain limited, however thanks to developments in research surrounding generative artificial intelligence the combination of both of these desires may be made possible through the use of the latest Generative Adversarial Networks. The aim of this project is to implement and compare four different clustering methods. These methods will be used to generate classification labels from gameplay images which will then be given as input to a generative network to create photorealistic equivalents. It will then be determined which method is most suitable for this task by comparing their initial classification performance and the results from the photorealistic images they are used to generate. In order to compare classification performance, the Dice coefficient was calculated for each classification image generated, using a ground truth image to represent perfect segmentation. It was found that good classification performance does not necessarily lead to superior GauGAN output images, and overall the best performing method for this task was Region-Growing due to the spatial consideration in its approach.