3D Point Cloud Analysis: Traditional, Deep Learning, and Explainable Machine Learning Methods

✍ Scribed by Shan Liu, Min Zhang, Pranav Kadam, C.-C. Jay Kuo

• Publisher: Springer
• Year: 2021
• Tongue: English
• Leaves: 156
• Edition: 1
• Category: Library

✦ Synopsis

This book introduces the point cloud; its applications in industry, and the most frequently used datasets. It mainly focuses on three computer vision tasks -- point cloud classification, segmentation, and registration -- which are fundamental to any point cloud-based system. An overview of traditional point cloud processing methods helps readers build background knowledge quickly, while the deep learning on point clouds methods include comprehensive analysis of the breakthroughs from the past few years. Brand-new explainable machine learning methods for point cloud learning, which are lightweight and easy to train, are then thoroughly introduced. Quantitative and qualitative performance evaluations are provided. The comparison and analysis between the three types of methods are given to help readers have a deeper understanding.

With the rich deep learning literature in 2D vision, a natural inclination for 3D vision researchers is to develop deep learning methods for point cloud processing. Deep learning on point clouds has gained popularity since 2017, and the number of conference papers in this area continue to increase. Unlike 2D images, point clouds do not have a specific order, which makes point cloud processing by deep learning quite challenging. In addition, due to the geometric nature of point clouds, traditional methods are still widely used in industry. Therefore, this book aims to make readers familiar with this area by providing comprehensive overview of the traditional methods and the state-of-the-art deep learning methods.

A major portion of this book focuses on explainable machine learning as a different approach to deep learning. The explainable machine learning methods offer a series of advantages over traditional methods and deep learning methods. This is a main highlight and novelty of the book. By tackling three research tasks -- 3D object recognition, segmentation, and registration using our methodology -- readers will have a sense of how to solve problems in a different way and can apply the frameworks to other 3D computer vision tasks, thus give them inspiration for their own future research. 

Numerous experiments, analysis and comparisons on three 3D computer vision tasks (object recognition, segmentation, detection and registration) are provided so that readers can learn how to solve difficult Computer Vision problems.

✦ Table of Contents

Preface
Contents
Author Biographies
1 Introduction
1.1 Introduction
1.2 3D Point Clouds
1.2.1 Point Cloud Formation
1.2.2 Comparison with Other Visual Data Forms
1.3 Point Cloud Processing
1.3.1 Registration
1.3.2 Classification
1.3.3 Semantic Segmentation
1.3.4 Odometry
1.4 Applications
1.5 Datasets
1.5.1 ModelNet40
1.5.2 ShapeNet
1.5.3 S3DIS
1.5.4 3D Match
1.5.5 KITTI
1.6 Summary
References
2 Traditional Point Cloud Analysis
2.1 Filtering
2.1.1 Downsampling
2.1.2 Noise Removal
2.2 Nearest Neighbor Search
2.2.1 Binary Search Tree
2.2.2 k-Dimensional Tree
2.2.3 Octree
2.3 Model Fitting
2.3.1 Least Squares Fitting
2.3.2 Hough Transform
2.3.3 Random Sample Consensus
2.4 Point Cloud Features
2.4.1 Feature Detectors
2.4.1.1 Harris 3D/6D
2.4.1.2 Intrinsic Shape Signatures
2.4.2 Feature Descriptors
2.4.2.1 Point Feature Histogram
2.4.2.2 Fast Point Feature Histogram
2.4.2.3 Signature of Histograms of Orientations
2.5 Classification and Segmentation
2.6 Registration
2.6.1 Iterative Closest Point (ICP)
2.6.2 Point-to-Plane ICP
2.6.3 Generalized ICP
2.6.4 Global Registration
References
3 Deep Learning-Based Point Cloud Analysis
3.1 Introduction
3.2 Classification and Segmentation
3.2.1 PointNet
3.2.2 PointNet++
3.2.3 Dynamic Graph CNN
3.2.4 PointCNN
3.2.5 PointSIFT
3.2.6 Point Transformer
3.2.7 RandLA-Net
3.3 Registration
3.3.1 PointNetLK
3.3.2 Deep Closest Point
3.3.3 PRNet
3.3.4 3D Match
3.3.5 PPFNet
3.3.6 Deep Global Registration
References
4 Explainable Machine Learning Methods for Point Cloud Analysis
4.1 Successive Subspace Learning on 2D Images
4.1.1 Data-Driven Saak Transform
4.1.2 Handwritten Digit Recognition by Saak Transform
4.1.3 Interpretable Convolutional Neural Networks via Feedforward Design
4.1.4 PixelHop
4.1.5 PixelHop++
4.2 Classification and Part Segmentation
4.2.1 PointHop
4.2.2 PointHop++
4.2.3 Unsupervised Feedforward Feature (UFF)
4.3 Registration
4.3.1 Salient Points Analysis (SPA)
4.3.2 R-PointHop
4.4 Other Applications of Successive Subspace Learning
4.4.1 FaceHop
4.4.2 DefakeHop
4.4.3 AnomalyHop
References
5 Conclusion and Future Work
5.1 Conclusion
5.2 Future Work
Index