Advancing 6DoF Object Pose Estimation: Keypoint Voting, Optimal Keypoint Sampling, and Bridging the Simulation-to-real Gap

This thesis explores advancements in six-degree-of-freedom (6DoF) object pose estimation using point cloud and RGB-D data. Three novel methodologies are proposed to address distinct challenges in this field. First, RCVPose3D introduces a cascaded keypoint voting framework that separates semantic segmentation from keypoint regression, incorporating pairwise constraints and a Voter Confident Score to improve accuracy. RCVPose3D achieves state-of-the-art results on Occlusion LINEMOD (74.5%) and YCB-Video (96.9%), outperforming traditional RGB and RGB-D methods. Second, KeyGNet leverages a graph network to optimize the keypoint selection, improving accuracy and efficiency by learning dispersed, evenly distributed keypoints. KeyGNet enhances performance across all metrics, notably increasing ADD(S) on Occlusion LINEMOD by 16.4% and closing the single-to-multiobject training gap. Finally, RKHSPose addresses the simulation-to-real domain gap through a self-supervised framework using a learnable kernel in RKHS and an adapter network pre-trained on synthetic data. This approach achieves competitive results against fully supervised methods, requiring no real groundtruth annotations. Together, these contributions advance 6DoF pose estimation in accuracy, efficiency, and adaptability across diverse datasets and scenarios.

Department dc:contributor.department

Electrical and Computer Engineering

Year dc:date.issued

2025

Author dc:creator

• Wu, Yangzheng

Advisor dc:contributor.supervisor

• Greenspan, Michael

• 6DoF pose

Statement dc:rights

• Attribution 4.0 International

Licence dc:rights.uri

• http://creativecommons.org/licenses/by/4.0/

Language dc:language.iso

eng

Handle dc:identifier.uri

https://hdl.handle.net/1974/34297

OAI identifier oai:identifier

oai:queensu.scholaris.ca:1974/34297