Automatically generating textured buildings using reconstructive and statistical methods

Abstract

Realistic 3D models of cities are extremely important for generating synthetic datasets for deep learning models. Current algorithms generate models of cities that are either monotonous, unrealistic, or must be specified with many procedural rules. In this thesis we present AutoCity, a system for generating realistic textured buildings given only the footprints of the buildings. AutoCity generates buildings using reconstructive and procedural methods, and then textures the buildings using a known adversarial neural network, FrankenGAN. We provide images of our output and extensive documentation on how the system can be run and extended.