Terrain Edge Stitching Based On Least Squares Generative Adversarial Networks

Virtual terrain is widely used in computer graphics and game, etc. This paper proposes an end-to-end deep learning approach for a 3D unit terrain generation from a simple sketch of terrain features using Least Squares Generative Adversarial Networks (LSGANs). Then edge regeneration strategy is used to seamlessly stitches the edges of multiple unit terrains. Unlike the existing works, edge regeneration strategy regenerates the connection area between multiple unit terrains, to expand the terrain size and shape while the existing terrain unchanged. LSGANs trained by using real-world terrains and their sketched counterparts. Experimental results show that the proposed approach produces a vast variety of complex terrains in high level of realism without fixed shapes and sizes only for a minimum sketch cost.

1. Introduction

Terrain generation is an indispensable visual content when creating realistic representations of virtual worlds. In computer games field, the size of terrain does not need to be fixed at the beginning, which saves memory overhead and the designer's workload. In order to enhance the user's interaction with the virtual worlds, users are even allowed to create their own terrain. Existing works often completely regenerate the terrain when expanding its size, instead of only generating the extension part while the existing part unchanged.