Interactivity, Game Creation, Design, Learning, and Innovation. 6th International Conference, ArtsIT 2017, and Second International Conference, DLI 2017, Heraklion, Crete, Greece, October 30–31, 2017, Proceedings

Research Article

Deep Convolutional Generative Adversarial Network for Procedural 3D Landscape Generation Based on DEM

Download
328 downloads
  • @INPROCEEDINGS{10.1007/978-3-319-76908-0_9,
        author={Andreas Wulff-Jensen and Niclas Rant and Tobias M\`{u}ller and Jonas Billeskov},
        title={Deep Convolutional Generative Adversarial Network for Procedural 3D Landscape Generation Based on DEM},
        proceedings={Interactivity, Game Creation, Design, Learning, and Innovation. 6th International Conference, ArtsIT 2017, and Second International Conference, DLI 2017, Heraklion, Crete, Greece, October 30--31, 2017, Proceedings},
        proceedings_a={ARTSIT \& DLI},
        year={2018},
        month={3},
        keywords={GAN Deep Convolutional Generative Adversarial Network PCG Procedural generated landscapes Digital Elevation Maps (DEM) Heightmaps Games 3D landscapes},
        doi={10.1007/978-3-319-76908-0_9}
    }
    
  • Andreas Wulff-Jensen
    Niclas Rant
    Tobias Møller
    Jonas Billeskov
    Year: 2018
    Deep Convolutional Generative Adversarial Network for Procedural 3D Landscape Generation Based on DEM
    ARTSIT & DLI
    Springer
    DOI: 10.1007/978-3-319-76908-0_9
Andreas Wulff-Jensen1,*, Niclas Rant1,*, Tobias Møller1,*, Jonas Billeskov1,*
  • 1: Aalborg University Copenhagen
*Contact email: awj@create.aau.dk, nrant14@student.aau.dk, tnma14@student.aau.dk, jbille14@student.aau.dk

Abstract

This paper proposes a novel framework for improving procedural generation of 3D landscapes using machine learning. We utilized a Deep Convolutional Generative Adversarial Network (DC-GAN) to generate heightmaps. The network was trained on a dataset consisting of Digital Elevation Maps (DEM) of the alps. During map generation, the batch size and learning rate were optimized for the most efficient and satisfying map production. The diversity of the final output was tested against Perlin noise using Mean Square Error [1] and Structure Similarity Index [2]. Perlin noise is especially interesting as it has been used to generate game maps in previous productions [3, 4]. The diversity test showed the generated maps had a significantly greater diversity than the Perlin noise maps. Afterwards the heightmaps was converted to 3D maps in Unity3D. The 3D maps’ perceived realism and videogame usability was pilot tested, showing a promising future for DC-GAN generated 3D landscapes.