Abstract:The 3D point cloud (3DPC) has significantly evolved and benefited from the advance of deep learning (DL). However, the latter faces various issues, including the lack of data or annotated data, the existence of a significant gap between training data and test data, and the requirement for high computational resources. To that end, deep transfer learning (DTL), which decreases dependency and costs by utilizing knowledge gained from a source data/task in training a target data/task, has been widely investigated. Numerous DTL frameworks have been suggested for aligning point clouds obtained from several scans of the same scene. Additionally, DA, which is a subset of DTL, has been modified to enhance the point cloud data's quality by dealing with noise and missing points. Ultimately, fine-tuning and DA approaches have demonstrated their effectiveness in addressing the distinct difficulties inherent in point cloud data. This paper presents the first review shedding light on this aspect. it provides a comprehensive overview of the latest techniques for understanding 3DPC using DTL and domain adaptation (DA). Accordingly, DTL's background is first presented along with the datasets and evaluation metrics. A well-defined taxonomy is introduced, and detailed comparisons are presented, considering different aspects such as different knowledge transfer strategies, and performance. The paper covers various applications, such as 3DPC object detection, semantic labeling, segmentation, classification, registration, downsampling/upsampling, and denoising. Furthermore, the article discusses the advantages and limitations of the presented frameworks, identifies open challenges, and suggests potential research directions.

What problem does this paper attempt to address?

The problems that this paper attempts to solve mainly focus on the challenges faced in 3D point cloud (3DPC) processing. Specifically, these problems include: 1. **Insufficient data or insufficient labeled data**: In 3D point cloud processing, high - quality labeled data is usually very scarce, which limits the training effect of deep - learning models. 2. **Significant gap between training data and test data**: Due to the different collection environments and conditions of 3D point cloud data, there may be a large distribution difference between training data and test data, resulting in insufficient generalization ability of the model. 3. **High demand for computing resources**: Processing large - scale 3D point cloud data requires a large amount of computing resources, which is an important bottleneck in practical applications. To address these challenges, the paper explores the application of deep transfer learning (DTL). DTL trains the target data/task by using the knowledge obtained from the source data/task, thereby reducing the dependence on a large amount of labeled data and improving the generalization ability of the model. Specifically, the main contributions of the paper include: - **Review of the background**: Introduces the basic concepts of DTL, related datasets, and evaluation metrics. - **Classification and comparison**: Proposes a detailed classification system and systematically classifies and compares existing DTL methods. - **Application scenarios**: Covers a variety of applications such as 3D point cloud object detection, semantic labeling, segmentation, classification, registration, down - sampling/up - sampling, and denoising. - **Advantages and limitations**: Discusses the advantages and limitations of existing frameworks, identifies open challenges, and proposes future research directions. Through these contents, the paper aims to provide a comprehensive review for 3D point cloud understanding, helping researchers and practitioners better understand and apply deep transfer learning techniques.

Advancing 3D Point Cloud Understanding through Deep Transfer Learning: A Comprehensive Survey

Recent Advancements in Learning Algorithms for Point Clouds: An Updated Overview

Deep Learning for 3D Point Cloud Understanding: A Survey

Deep Learning for 3D Point Clouds: A Survey

Review: Deep Learning on 3D Point Clouds

Deep learning based computer vision under the prism of 3D point clouds: a systematic review

Recent Advances and Perspectives in Deep Learning Techniques for 3D Point Cloud Data Processing

3D Point Cloud for Objects and Scenes Classification, Recognition, Segmentation, and Reconstruction: A Review

Deep Learning for LiDAR Point Cloud Classification in Remote Sensing

A Survey on Deep Learning Based Segmentation, Detection and Classification for 3D Point Clouds

Deep Learning for 3D Point Cloud Enhancement: A Survey

Deep Learning-based 3D Point Cloud Classification: A Systematic Survey and Outlook

Advancements in Point Cloud-Based 3D Defect Detection and Classification for Industrial Systems: A Comprehensive Survey

Comprehensive Review of Deep Learning-Based 3D Point Cloud Completion Processing and Analysis

UNDERSTANDING 3D POINT CLOUD DEEP NEURAL NETWORKS BY VISUALIZATION TECHNIQUES

Deep Learning for LiDAR-only and LiDAR-fusion 3D Perception: a Survey

Deep learning for 3D object recognition: A survey

A Review of Deep Transfer Learning and Recent Advancements

Deep Learning-Based 3D Instance and Semantic Segmentation: A Review