SGRTmreg: A Learning-Based Optimization Framework for Multiple Pairwise Registrations.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Data Processing

Background:

• Point cloud registration is crucial for 3D reconstruction and object tracking.
• Existing deep learning and learning-based optimization methods have distinct advantages but also limitations in pairwise registration.
• There is a need for methods that combine robustness, stability, and efficiency for multi-instance registration.

Purpose of the Study:

• To propose a novel computational framework, SGRTmreg, for efficient and robust multi-instance point cloud registration.
• To leverage the strengths of both deep learning and learning-based optimization for improved registration performance.
• To achieve accurate, stable, and less time-consuming registration for multiple point cloud instances.

Main Methods:

• The SGRTmreg framework integrates a Searching scheme, Graph-based Reweighted discriminative optimization (GRDO), and a Transfer module.
• The Searching scheme identifies the most relevant point cloud from a collection for registration.
• GRDO learns alignment regressors, and the Transfer module applies these to achieve multi-instance registration to a target point cloud.

Main Results:

• SGRTmreg demonstrated superior accuracy, robustness, and stability in extensive registration experiments.
• The framework successfully registered multiple point clouds to a target point cloud using shared regressors.
• Experimental results show SGRTmreg outperforms state-of-the-art deep learning and traditional registration methods.

Conclusions:

• SGRTmreg offers a powerful solution for multi-instance point cloud registration.
• The proposed framework effectively combines learning-based optimization and deep learning principles.
• SGRTmreg achieves state-of-the-art performance in accuracy, robustness, and stability for point cloud registration tasks.