Abstract:Learning-based point cloud registration methods can handle clean point clouds well, while it is still challenging to generalize to noisy, partial, and density-varying point clouds. To this end, we propose a novel point cloud registration framework for these imperfect point clouds. By introducing a neural implicit representation, we replace the problem of rigid registration between point clouds with a registration problem between the point cloud and the neural implicit function. We then propose to alternately optimize the implicit function and the registration between the implicit function and point cloud. In this way, point cloud registration can be performed in a coarse-to-fine manner. By fully capitalizing on the capabilities of the neural implicit function without computing point correspondences, our method showcases remarkable robustness in the face of challenges such as noise, incompleteness, and density changes of point clouds.

What problem does this paper attempt to address?

This paper attempts to address several key challenges in Point Cloud Registration, especially for point cloud data with noise, partial missing, and uneven density. Traditional methods perform poorly when dealing with these issues. For example, correspondence-based methods are easily affected by noise and density changes, while direct methods can handle clean data but have limited performance in complex scenarios. To overcome these challenges, the paper proposes a new point cloud registration framework **SDFReg**, which transforms the rigid registration problem into a registration problem between the point cloud and the neural implicit surface by introducing Neural Implicit Representation. Specifically, SDFReg uses the Signed Distance Function (SDF) to represent the target point cloud and estimates the transformation parameters by optimizing the distance metric between the source point cloud and the neural implicit surface. This method avoids computing point correspondences, thereby improving robustness to noise, incompleteness, and density variations. ### Main Contributions: 1. **Proposed SDFReg**: A new point cloud registration framework that fully leverages the capabilities of neural implicit functions, eliminating the need to find correspondences. 2. **Introduced a coarse-to-fine learning strategy**: Alternately optimizing implicit function fitting and registration, further improving registration quality. 3. **Demonstrated robustness to point cloud defects**: Capable of performing well on point cloud data with noise, partial missing, and density variations, and can generalize to different categories. ### Method Overview: 1. **Problem Definition**: The goal of 3D rigid point cloud registration is to find a rigid transformation matrix that aligns the source point cloud to the target point cloud. 2. **Using SDF for Registration**: Represent the target point cloud as a neural implicit surface through SDF, then optimize the distance metric between the source point cloud and the implicit surface. 3. **Learning Neural SDF**: Use a Multi-Layer Perceptron (MLP) to regress a continuous SDF and train it through a self-supervised loss function. 4. **Coarse-to-Fine Registration Strategy**: Gradually optimize the SDF and registration by dynamically acquiring query points at different distances, improving the registration accuracy in local areas. ### Experimental Results: - Experiments on the **ModelNet40** dataset show that SDFReg performs well on both clean data and Gaussian noise data. - SDFReg also outperforms other methods on partially visible point clouds and point clouds with density variations. - Experiments on the **3DMatch** dataset further validate the robustness of SDFReg. In summary, SDFReg effectively addresses multiple challenges in point cloud registration by introducing neural implicit representation and a coarse-to-fine learning strategy, particularly excelling in handling point cloud data with noise, partial missing, and density variations.

SDFReg: Learning Signed Distance Functions for Point Cloud Registration

Learning-based Point Cloud Registration: A Short Review and Evaluation.

DeepICP: An End-to-End Deep Neural Network for 3D Point Cloud Registration

PointDifformer: Robust Point Cloud Registration With Neural Diffusion and Transformer

Deep Learning Based Point Cloud Registration: an Overview

Deep learning-based point cloud registration: a comprehensive investigation

A Novel and Efficient Point Cloud Registration by Using Coarse-to-Fine Strategy Integrating PointNet

An Automatic and Robust Point Cloud Registration Framework Based on View-Invariant Local Feature Descriptors and Transformation Consistency Verification

Point cloud registration for LiDAR and photogrammetric data: A critical synthesis and performance analysis on classic and deep learning algorithms

Reg-NF: Efficient Registration of Implicit Surfaces within Neural Fields

SDRSAC: Semidefinite-Based Randomized Approach for Robust Point Cloud Registration Without Correspondences

Point Cloud Registration Based on Direct Deep Features With Applications in Intelligent Vehicles

GenReg: Deep Generative Method for Fast Point Cloud Registration

Partial Point Cloud Registration with Deep Local Feature

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

End-to-End 3D Point Cloud Learning for Registration Task Using Virtual Correspondences.

Robust-DefReg: A Robust Deformable Point Cloud Registration Method based on Graph Convolutional Neural Networks

GraphReg: Dynamical Point Cloud Registration With Geometry-Aware Graph Signal Processing

Robust Point Cloud Registration Using Geometric Spatial Refinement

3D Point Cloud Registration Method Based on Deep and Shallow Features Combined with Attention Enhancement