Machine Learning requires data. Without the availability of large, high-quality datasets, the success of deep learning in recent years would not have been possible. Data is the fundamental building block in developing AI pipelines. However, due to the limitations in measurement tools, lack of control and immutability of real-life datasets, the general approach to developing machine learning solutions has evolved to be model-centric. This Dissertation explores the possibility of Data-centric AI by looking at the development of a novel technology | flexible photorealistic simulations | that can generate labeled datasets for use in lieu of real data in various fields of deep-learning accelerated computer vision. In each chapter of this work, we'll follow a major phase shift that represents a forward step in the applications of this field. From proof of concept, Improving existing methods, Applications on hard tasks, to achieving state-of-the-art performance.