Online Evaluation of Depth-Based 6D Object Pose Estimation

Abstract

In this thesis, we demonstrate the effectiveness of deep learning in a subtask of autonomous geometric tasks requiring human-level and above performance, such as operating complex machinery like driving and excavation, and robotic grasping. A key challenge in these tasks is accurately estimating the 6D pose (i.e., 3D rotation and 3D translation) of an observed object. To address this challenge, we present an online pose estimation pipeline that uses a depth and template-based model, which leverages a pre-existing 3D model of the object, along with an offthe-shelf mobile iOS tool for 3D reconstruction from a sequence of RGB images. We evaluate the performance of the pipeline on several example objects and analyze both successful and unsuccessful cases, highlighting the diffculties that can arise in pose estimation. Future work could include incorporating object detection into the pose estimation architecture and using more expressive texture representations to handle refective surfaces.