Provable and Control-Theoretic Methods for Deep Object Pose Estimation

Abstract

In this dissertation, we consider the task of object pose estimation using deep neural networks. We draw our motivation from the fact that neural networks have shown to be successful at the task of pose estimation, but are poorly theoretically understood and lack meaningful performance guarantees. As a result, our aim in this dissertation is to analyze pose estimation neural networks by developing provable performance guarantees, as well as connecting pose estimation to control theory. We take four different approaches in our analysis. First, we consider object pose estimation from the standpoint of observability in control theory, using the observability Gramian as our main tool for analysis. Next, we explore the idea of estimating the pose of a dynamic object by applying an unscented filter to pose estimates from a neural network. Next, we derive analytical bounds on the local Lipschitz constants of neural networks with ReLU activations. Finally, we consider the task of developing sensitivity bounds for pose estimation neural networks, and construct a pose estimation network with provable bounds for both the rotation and position estimates.