Deep Learning for 3D Point Clouds

Abstract

The last decade has witnessed the great success of deep learning theories, methods, and applications in almost all science and engineering fields.As is implied by the name, deep learning leverages the powerful capability of deep neural networks as machine learning models to fulfill complex prediction, understanding, and decision problems, as long as there are large-scale datasets and sufficient computing power.For computer vision tasks, people are now not satisfied with 2D images any more, and in these circumstances, 3D modeling capability from 3D point clouds becomes much more important and popular.For 3D human and machine perception, 3D point clouds can provide the immersive visual experience and the high-precision 3D modeling for 3D objects, indoor and outdoor scenes.Moreover, recently large language model (LLM) and Multi-modal LLM have been extensively investigated, and 3D pre-trained models and 3D large models are expected to bring new opportunities to reshape the world, especially by the means of embodied AI.Consisting of 11 chapters, this book focuses on the deep learning-based point cloud technologies, and seeks to provide readers with an in-depth understanding of point cloud processing methods in a textbook manner, including enhancement, analysis, pre-trained models and large models, multi-modal large models, open source projects, and engineering applications, etc.This book puts an emphasis on the perspectives of deep learning, 3D human and machine perception, and large models.The detailed chapters are organized as follows:Chapter 1 presents an overview of the 3D world representation with point clouds, including representative datasets, processing tasks, and applications.Chapter 2 introduces the fundamental background knowledge of deep learning, and several basic deep neural networks for point cloud tasks.Chapters 3 and 4 demonstrate the deep learning-based point cloud enhancement principles and methods, including upsampling, downsampling, frame interpolation, completion, and denoising.Chapters 5 and 6 delve into the deep learning-based point cloud analysis principles and methods, including classification and segmentation, object detection, tracking, retrieval, registration, and multimodal analysis.v vi

Topics & Concepts