Pseudo Label Learning for Partial Point Cloud Registration.

Area of Science:

• Computer Vision
• 3D Data Processing

Background:

• Partial point cloud registration is vital for 3D applications but often requires ground truth labels.
• Existing methods struggle with accuracy and efficiency due to reliance on manual or imprecise automatic labeling.

Purpose of the Study:

• To develop an unsupervised partial point cloud registration method that eliminates the need for ground truth labels.
• To introduce a novel framework, PSEudo Label learning (PSEL), for accurate and efficient partial point cloud registration.

Main Methods:

• PSEL utilizes complementary tasks to generate pseudo labels for overlap regions and correspondences.
• The method employs an overlap estimation module and a correspondence filtering module, supervised by generated pseudo labels.
• A two-pipeline correspondence filtering module uses feature similarity and differences for robust correspondence identification.

Main Results:

• PSEL achieves effective unsupervised partial point cloud registration across synthetic (ModelNet40) and real-world datasets (3DMatch, KITTI).
• The method demonstrates improved accuracy in estimating overlap regions and correspondences without ground truth supervision.
• Experimental validation confirms the robustness and effectiveness of the PSEL framework.

Conclusions:

• PSEL offers a viable unsupervised solution for partial point cloud registration, overcoming the limitations of label-dependent methods.
• The proposed pseudo-labeling strategy enhances the learning of critical registration components.
• This work contributes to advancing 3D computer vision applications requiring precise point cloud alignment.