Recovering 6D object pose and semantic information from RGBD image

Abstract

The applications of 6-Dimentional (6D) pose estimation plays an important role in today’s technology. The increasing prominence of robotics, automation, and augmented reality necessitates the need for precise and effective 6D pose information to ensure its successful execution. For industrial applications, easily replicable datasets and robust pose estimations are crucial in determining the possible applications for the 6D pose method. The use of color (Red, Green, Blue) and depth information, also known as RGBD images have emerged as a popular source of data for 6D pose estimation. This project explores two popular RGBD pointcloud processing methods for 6D pose estimation: Normalized Object Coordinate Space (NOCS) and Deep Point-wise 3D Keypoints Voting Network (PVN3D). A review of the methodology, results, and dataset requirements regarding the two methods were discussed. Upon further inspecting and comparison of the methods, the replication of the NOCS dataset was found difficult to replicate with the limited knowledge and resources in hand, while PVN3D was found to meet the requirements. The LineMOD dataset which was utilized by PVN3D was shown to be replaceable using the in-house equipment provided. Furthermore, the results obtained from the training and testing of NOCS did not perform up to expectations, while the PVN3D results were acceptable. Overall, this project concludes that PVN3D is a viable 6D pose estimation method for industrial applications.