Match Normalization: Learning-Based Point Cloud Registration for 6D Object Pose Estimation in the Real World.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Learning-based 3D object registration methods excel on synthetic data but fail on real-world data.
• Key challenges include sensitivity to rotation range and differing feature distributions between point clouds.

Purpose of the Study:

• To investigate and address the limitations of current learning-based 6D pose estimation from point clouds.
• To develop generalizable strategies applicable to existing 3D object registration frameworks.

Main Methods:

• Introduced a directly supervised loss function to overcome SVD-based loss limitations with rotation.
• Developed a new normalization strategy, Match Normalization, to handle feature distribution differences.

Main Results:

• Implemented contributions in DCP and IDAM frameworks, demonstrating significant improvements.
• Achieved meaningful results on real-world datasets (TUD-L, LINEMOD, Occluded-LINEMOD) for the first time with learning-based methods.

Conclusions:

• The proposed loss function and normalization strategy are effective for real-world 6D pose estimation.
• These advancements are expected to drive future developments in point cloud registration.