Title: Leveraging Simulation for Computer Vision

Abstract:

A large amount of labeled data is required to train deep neural networks. The process of data annotation on such a large scale is expensive and time-consuming. A promising alternative in this regard is to use simulation to generate labeled synthetic data. However, a network trained solely on synthetic data may not generalize to real data due to domain gap. This thesis is a study of domain randomization (DR) as a solution to reduce this domain gap. We provide theoretical insights into the effectiveness of DR supported with empirical results for image classification, object detection and pose estimation. Finally, our analysis highlights key limitations of the approach which we alleviate by proposing adversarial domain randomization (ADR), a more data efficient variant of DR.

Committee:

Kris Kitani (Chair)

Deva Ramanan

David Held

Lerrel Pinto