Deep Convolutional GANs for Real-Time Procedural Terrain Generation Systems
Year: 2019
Type: dissertation
Abstract: Many animated films, computer games, and simulations attempt to mimic \nnaturalistic terrain in a three-dimensional environment in hope of immersing the audience \ninto the world it creates. Creating such terrain in a three-dimensional environment is not a \ntrivial task, yet many methods have been introduced to solve this problem. Hand\nsculpting solutions require environmental artists to manually design the terrain with \nsoftware that allows them to pull, carve, and model terrain in real-time. Although a valid \napproach, it can take a lot of time and effort to achieve believable terrain. Additionally, it \nrequires a high level of expertise in terrain modeling software, something that not many \nof those who wish to create realistic terrain poses. To alleviate these barriers, methods \nthat allow the computer to automatically generate terrain have been invented. These \nmethods involve creating specialized algorithms to procedurally generate terrain. These procedural generation systems are good at producing naturalistic-looking terrain, yet like \nthe hand-sculpting approach, designing these algorithms is no easy feat. A significant \namount of ingenuity and knowledge of geophysics is required to implement these \nalgorithms. Furthermore, each algorithm is specialized to a certain type of terrain, and a \nnew algorithm must be redesigned and implemented to support other types of terrains. \nTherefore, algorithmically producing terrains is not a flexible approach. The limitations \nof the suggested approaches make the terrain creation process difficult and time\nconsuming. \nIn this thesis, a new approach is presented for generating three-dimensional \nterrain by using deep learning techniques to simplify the terrain creation process. In this \napproach, a Deep Convolutional Generative Adversarial Network (DCGAN) is utilized to \ngenerate terrain heightmaps of common physical terrain features. These terrain \nheightmaps are then used to generate terrain in a three-dimensional environment with \nthese physical features. The benefit of this approach allows for the creation of naturalistic \nterrain without the need to hand-craft or implement an algorithm. This approach aims to \nmake the process of creating any type of terrain in three-dimensional environments more \nflexible than the current approaches. The thesis includes an analysis of the terrain \ngeneration performance of the DCGAN-based approach against the algorithmic-based \nterrain generation approach. A process is also introduced on how to set up a DCGAN\nbased procedural terrain generation system and the results of using such system is \nexamined. The thesis then concludes with a discussion on how the work in this thesis can \nfurther be extended into improving the DCGAN procedural terrain generation outputs and how a DCGAN can be used in other areas of computer graphics to simplify many of \nthe processes common in the computer graphics field.
Author: Lopez-Garcia, Edgar
Institution: California State University, Northridge
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7065012304
