HarSoNet: a two-stage point cloud registration method integrating soft and hard matching.

Area of Science:

• Computer Vision
• Robotics
• Geometric Computing

Background:

• Point cloud registration is crucial for 3D applications like SLAM and scene reconstruction.
• Existing coarse-to-fine methods struggle with outlier correspondences and sensitivity to initial matching.
• Accurate correspondence establishment in unordered point clouds remains a significant challenge.

Purpose of the Study:

• To introduce HarSoNet, a novel two-stage Hard-to-Soft Network for robust and accurate end-to-end point cloud registration.
• To address the limitations of hard matching and improve correspondence quality in registration pipelines.
• To enhance the generalization performance of point cloud registration algorithms.

Main Methods:

• HarSoNet employs a hybrid similarity fusion module for superpoint correspondence generation in the hard matching stage.
• Fuzzy patch correspondences are created by grouping superpoint correspondences and their neighbors.
• A soft matching stage refines patch correspondences into point correspondences by adjusting similarity matrices.

Main Results:

• HarSoNet achieved high registration accuracy with Error(R) = 1.376 and Error(t) = 0.015.
• The method demonstrated strong generalization performance on noisy, partially overlapping point clouds.
• The two-stage hard-to-soft approach effectively reduces outliers and improves correspondence robustness.

Conclusions:

• HarSoNet offers a significant advancement in point cloud registration by overcoming the limitations of traditional methods.
• The proposed network architecture provides a robust solution for accurate transformation parameter estimation.
• This work contributes to more reliable 3D reconstruction and simultaneous localization and mapping systems.