Abstract:Partial to Partial Point Cloud Registration (partial PCR) remains a challenging task, particularly when dealing with a low overlap rate. In comparison to the full-to-full registration task, we find that the objective of partial PCR is still not well-defined, indicating no metric can reliably identify the true transformation. We identify this as the most fundamental challenge in partial PCR tasks. In this paper, instead of directly seeking the optimal transformation, we propose a novel and general Sight View Constraint (SVC) to conclusively identify incorrect transformations, thereby enhancing the robustness of existing PCR methods. Extensive experiments validate the effectiveness of SVC on both indoor and outdoor scenes. On the challenging 3DLoMatch dataset, our approach increases the registration recall from 78\% to 82\%, achieving the state-of-the-art result. This research also highlights the significance of the decision version problem of partial PCR, which has the potential to provide novel insights into the partial PCR problem.

What problem does this paper attempt to address?

### Problems the paper attempts to solve The paper aims to solve the key challenges in the partial - to - partial point cloud registration (partial PCR) task. Specifically, when the overlap rate is low, the goal of the partial PCR task is still not clearly defined, resulting in no reliable metric to identify the correct transformation matrix. The author believes this is the most fundamental challenge in the partial PCR task. #### Main problem description 1. **Goal not clearly defined**: Compared with full - to - full and partial - to - full registration tasks, when the overlap rate is low, the goal of partial PCR is not yet clear. This makes it difficult for existing registration methods to reliably select the correct transformation. 2. **Lack of reliable metric**: In partial PCR, even if multiple possible transformation hypotheses are given, no known metric can reliably identify the correct transformation among them. Especially in the case of extremely low overlap rates, it is even uncertain whether such a metric exists. 3. **Limitations of existing methods**: Although some studies have proposed different descriptors and robust outlier rejection methods to deal with the problem of high outlier rates, these methods fail to solve the problem fundamentally. The author points out that the high outlier rate is a significant but not the most fundamental challenge. ### Proposed solutions To solve the above problems, the author proposes a new general constraint - Sight View Constraint (SVC). The core idea of SVC is to identify incorrect transformations by determining whether the transformed source point cloud blocks the line of sight between the target point cloud and the sensor, thereby enhancing the robustness of existing registration methods. #### Main contributions 1. **Proposing strict SVC constraints**: This constraint can significantly improve the robustness of existing registration methods by narrowing the transformation range by excluding obviously incorrect transformation hypotheses. 2. **Emphasizing the partial PCR decision - version problem**: The author analyzes the goal of the partial PCR task and emphasizes the importance of the decision - version problem. They divide the registration method into two stages: generating an effective set of hypotheses (including the correct hypothesis) and quickly identifying the correct transformation. 3. **Experimental verification**: Extensive experiments show that SVC can effectively improve the registration performance in both indoor and outdoor scenarios. In particular, on the challenging 3DLoMatch dataset, SVC increases the registration recall rate from 78% to 82%, reaching the state - of - the - art level. ### Summary By introducing the SVC constraint, this paper solves the key problems of unclear goals and lack of reliable metrics in the partial PCR task, significantly improving the robustness and accuracy of registration.

Sight View Constraint for Robust Point Cloud Registration

Efficient Single Correspondence Voting for Point Cloud Registration

An Effective Point Cloud Registration Method Based on Robust Removal of Outliers

Hierarchical Multiview Rigid Registration

SC^2-PCR: A Second Order Spatial Compatibility for Efficient and Robust Point Cloud Registration

Multi-view Non-rigid Registration Based on Multiple Thin-plate Splines

SGNet: Salient Geometric Network for Point Cloud Registration

SGFeat: Salient Geometric Feature for Point Cloud Registration

A new method for point cloud registration: Adaptive relation-oriented convolution and recurrent correspondence-walk

R-PCR: Recurrent Point Cloud Registration Using High-Order Markov Decision

Clique-like Point Cloud Registration: A Flexible Sampling Registration Method Based on Clique-like for Low-Overlapping Point Cloud

CentroidReg: A Global-to-Local Framework for Partial Point Cloud Registration.

Partial-to-Partial Point Cloud Registration by Rotation Invariant Features and Spatial Geometric Consistency

Hybrid Optimization with Unconstrained Variables on Partial Point Cloud Registration

Efficient and Robust Point Cloud Registration via Heuristics-guided Parameter Search

RGB-D Point Cloud Registration Based on Salient Object Detection

Learning and Matching Multi-View Descriptors for Registration of Point Clouds

A new method for two-stage partial-to-partial 3D point cloud registration: multi-level interaction perception

Dual-Branch Deep Point Cloud Registration Framework for Unconstrained Rotation

RGB-D Point Cloud Registration Via Infrared and Color Camera

Efficient and Deterministic Search Strategy Based on Residual Projections for Point Cloud Registration with Correspondences