PCKRF: Point Cloud Completion and Keypoint Refinement With Fusion Data for 6D Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Traditional point cloud registration methods like ICP struggle with high initial pose accuracy.
• Deep learning advancements necessitate specialized pose refinement techniques for 6D pose estimation.
• Existing methods lack specific design for pose refinement, limiting effectiveness.

Purpose of the Study:

• To introduce a novel pose refinement pipeline, Point Cloud Completion and Keypoint Refinement with Fusion Data (PCKRF), for 6D pose estimation.
• To enhance the accuracy and stability of 6D pose estimation, particularly in challenging scenarios.
• To provide a complementary method that integrates with existing 6D pose estimation techniques.

Main Methods:

• Developed a pose-sensitive point completion network utilizing local and global features with pose information.
• Proposed the Color supported Iterative KeyPoint (CIKP) method for point cloud registration, incorporating color information and keypoint-based registration for stability.
• Integrated the PCKRF pipeline with existing 6D pose estimation methods, such as the full flow bidirectional fusion network.

Main Results:

• The PCKRF pipeline demonstrated superior stability compared to existing approaches when optimizing high-precision initial poses.
• The method effectively complements most existing pose estimation techniques, leading to improved performance.
• Achieved promising results in challenging scenarios involving textureless and symmetrical objects.

Conclusions:

• The proposed PCKRF pipeline offers a robust solution for pose refinement in 6D pose estimation.
• PCKRF enhances the performance and stability of existing 6D pose estimation methods.
• The method shows significant potential for applications dealing with difficult object geometries and textures.