Assessing the Utility of GAN-Generated 3D Virtual Desert Terrain: A User-Centric Evaluation of Immersion and Realism

Terrain modeling is increasingly becoming an essential part of robotics and virtual reality (VR). This technology has the potential to transform how robots interact with the world and how we experience virtual reality. In virtual reality, terrain models are used to create immersive and realistic landscapes for users to explore and interact with. However, creating varied, demanding, and realistic terrains in simulation is challenging and time-consuming. This paper presents an automated system capable of generating synthetic terrain and investigates how well the generated virtual terrain suits VR systems. To automate, we leverage the generative adversarial networks (GAN) with post-enhancement to generate a diverse and immersive desert. We used an actual digital elevation model (DEM) (1 m resolution) of the Mojave Desert in California, USA, obtained from the USGS agency to train the generative model. We investigated by employing a head-mounted display (HMD) for a robust quality assessment, and the results indicate our system successfully generated virtual terrain with acceptable realism.