Robust 6-DoF Pose Estimation under Hybrid Constraints

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Two-stage pose estimation algorithms using heatmaps often lack accuracy and stability, especially with occluded objects.
• Accurate object pose estimation is critical for applications in robotics, augmented reality, and autonomous systems.
• Existing methods struggle with complex scenes and partial occlusions, limiting their real-world applicability.

Purpose of the Study:

• To develop a robust 6-DoF pose estimation algorithm that overcomes the limitations of current heatmap-based methods for occluded objects.
• To improve keypoint accuracy and overall pose estimation performance in challenging, cluttered environments.
• To enhance the stability and reliability of object pose estimation algorithms.

Main Methods:

• A novel loss function was formulated for improved heatmap regression quality and keypoint accuracy.
• The heatmap regression network was expanded with an added translation regression branch for enhanced pose constraints.
• A robust pose optimization module was employed to fuse heatmap and translation estimates, refining accuracy.

Main Results:

• The proposed algorithm achieved ADD(-S) accuracy rates of 93.5% on LINEMOD and 46.2% on Occlusion LINEMOD datasets.
• Demonstrated superior performance compared to state-of-the-art algorithms, significantly reducing mean estimation error.
• Showcased improved stability and maintained a maximum speed of 22 FPS, indicating efficiency.

Conclusions:

• The developed robust 6-DoF pose estimation algorithm offers significant improvements in accuracy and stability for occluded object pose estimation.
• The hybrid constraint approach effectively enhances performance in complex scenes, outperforming conventional methods.
• The algorithm presents a performant and efficient solution for real-time object pose estimation challenges.