A Semantic Segmentation System for generating context-based tile-maps

Abstract

Training a Generative Adversarial Network (GAN) involves a two-step process: training the generator network and training the discriminator network. The generator tries to generate realistic data, while the discriminator aims to distinguish between real and generated data. In this work we propose a semantic segmentation system that uses regular images for generating semantic maps through Tensor Flow framework. These maps are associated with a discrete set of tiles, which can be used for training generation of game style tile-maps. Besides the data-set creation, our solution also allows the creation of tile-maps based on image samples.