Single Voter Spreading for Efficient Correspondence Grouping and 3D Registration

Area of Science:

• Computer Vision
• 3D Data Processing
• Geometric Deep Learning

Background:

• Accurate 3D point cloud registration and recognition rely on consistent correspondences.
• Initial correspondences often suffer from outliers due to noise and limited overlap, challenging downstream tasks.
• Existing methods may struggle with efficiency and robustness in handling these noisy correspondences.

Purpose of the Study:

• To present a novel and efficient method for 3D correspondence grouping and registration.
• To address the challenge of outliers in initial point cloud correspondences.
• To achieve state-of-the-art performance with a computationally efficient and robust algorithm.

Main Methods:

• Introduced the Single Voter Spreading (SVOS) method.
• Leveraged low-order graph constraints within a single voter spreading voting scheme.
• Employed a two-stage voting process (single voter and spread voters voting) using only edge constraints.
• Utilized top-scored correspondences for robust transformation estimation.

Main Results:

• SVOS achieves new state-of-the-art performance in correspondence grouping and 3D registration.
• Demonstrated superior efficiency and robustness on benchmark datasets (U3M, 3DMatch/3DLoMatch, ETH, KITTI-LC).
• The method is light-weight and robust to graph construction parameters.

Conclusions:

• The proposed SVOS method offers an efficient and effective solution for 3D correspondence grouping and registration.
• SVOS successfully handles outliers and achieves high accuracy with reduced computational cost.
• This approach advances the field of 3D computer vision tasks requiring precise point cloud alignment.