3D Pose Estimation of Custom Objects Using Synthetic Datasets

Abstract

3D/6D pose estimation is a novel research area, part of the larger robotics sensing domain, focused on extracting 3D position and 3D rotation information using affordable hardware such as RGB-D or Stereoscopic depth cameras. Most estimators rely internally on a machine learning model for either the object detection phase or the entire 6D pose estimation loop. Thus, a custom machine learning (ML) model and dataset must be constructed and trained respectively in order to achieve the stated goal. The majority of the 3D/6D pose estimation models focus on standardized datasets so a custom dataset must also be created for each model. This article explores the benefits and challenges of artificially generated datasets on one 3D pose estimation model and the ML model transfer learning process. An accuracy test is conducted using real hardware.