Procedural synthetic data for self-driving cars using 3D graphics

Abstract

In this thesis we present CoSy, a configurable system for procedurally generating synthetic data for self-driving vehicles. To address the problem of data hungry vision-based learning algorithms used in self-driving vehicles, we develop a system that generates synthetic images, including class level annotations, of street scenes. To give researchers control over how data is generated, our system is designed to be configurable and extendable. We provide two example datasets generated by our system, and provide extensive documentation on how the system is architected, and how it can be extended.