3D Registration in 30 Years: A Survey

This is the official repository of 3D Registration in 30 Years：ASurvey (IEEE TPAMI), a comprehensive survey of recent progress in 3D registration for point clouds. For details, please refer to:

3D Registration in 30 Years: A Survey

1 Introduction

3D point clouds to a unified coordinate system, known as 3D point cloud registration, is a fundamental problem in numerous areas such as computer vision, computer graphics, robotics, and remote sensing.

2 Background

2.1 Basic Concepts

2.2 Datasets

2.3 Metrics

3 Pairwise Registration

3.1 Introduction

3.2 Pairwise Coarse Registration

3.2.1 Geometric methods

In correspondence-based methods, a crucial step is the generation of correspondences, which plays a key role in determining the accuracy and robustness of the registration process.

3.2.1.1 Correspondence-based Methods

3.2.1.1.1 Correspondence Generation

(1) Keypoint detection.

(2) Descriptors.

(3) Matching technique.

3.2.1.1.2 Correspondence Optimization

(1) Voting-based methods.

(2) Voting-free methods.

3.2.1.1.3 Transformation Estimation

(1) Sample-based methods.

(2) Parameter searching-based methods.

3.2.1.2 Correspondence-free Methods

3.2.2 Deep-learning-based methods

3.2.2.1 Supervised Methods

3.2.2.1.1 Non End-To-End

(1) Descriptor

(2) Others

3.2.2.1.2 End to End

(1) PointNet-based

(2) Graph-based

(3) ConvNet-based

(4) Transformer-based

3.2.2.2 Unsupervised Methods

(1) Self Reconstruction

(2) Mutual Reconstruction

(3) Metric Learning

(4) Regstration Prior

(5) Iteratively Registration

3.3 Pairwise Fine Registration

3.3.1 ICP Based

3.3.1.1 Sampling-aware

3.3.1.2 Matching-aware

3.3.1.3 Error metric-aware

3.3.2 GMM based

3.3.3 Others

4 Multi-view Registration

4.1 Multi-view Coarse Registration

Multi-view coarse registration aims to align point clouds from multiple views to form a coherent global model.

4.1.1 Geometric methods

(1) Corresspondence

(2) connected graph

4.1.2 Learning-based methods

(1)End-to-end

(2)Feature Descriptors

(3)Self-Supervised and Unsupervised

(4)Others

Multi-view fine registration

4.2 Multi-view Fine Registration

4.2.1 Point-based methods

4.2.2 Transformation-based methods

4.2.3 Optimization-based methods

5 Other Registration Problems

5.1 Cross-scale Registration

5.1.1 Traditional ICP-like methods

5.1.2 Deep-learning methods

5.2 Cross-source Registration

5.2.1 Conventional Optimization methods

5.2.2 Deep Neural Network methods

5.3 Color Point Cloud Registration

5.4 Multi-instance Point Cloud Registration

5.4.1 Multi-model Fitting methods

5.4.2 PPF-like methods

5.5 Sim2Real Registration