Instance-level 6D pose estimation based on multi-task parameter sharing for robotic grasping.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Six-dimensional pose estimation is crucial for robotics and computer vision tasks.
• Existing methods often struggle with accuracy due to object shape and size variations.
• Deep neural networks typically regress poses directly or via keypoint methods, leading to deviations.

Purpose of the Study:

• To enhance the accuracy and robustness of six-dimensional pose estimation.
• To address prediction deviations caused by object surface prominence and scale.
• To incorporate object category information into pose estimation networks.

Main Methods:

• Proposed a multi-task parameter sharing (PMP) framework for six-dimensional pose estimation.
• Extracted and fused image and point cloud features.
• Utilized an object classification auxiliary task to share knowledge and improve keypoint confidence.
• Employed an iterative optimization network for final pose refinement.

Main Results:

• Achieved improved accuracy in six-dimensional pose estimation on the LineMOD dataset.
• Reduced the accuracy gap for objects with different shapes.
• Demonstrated effective real-time pose estimation for robotic grasping experiments on a UR5 arm.

Conclusions:

• The PMP framework effectively integrates object category information for better pose estimation.
• The method shows significant improvements in accuracy and robustness across diverse object types.
• The approach is suitable for real-time applications, such as robotic grasping.