【论文阅读】Visual place recognition: A survey from deep learning perspective
Visual place recognition: A survey from deep learning perspective
PDF
[1] 张翔, 王磊, 孙杨. 视觉位置识别: 从深度学习视角展开的研究综述[J]. 模式识别, 2020.
公共数据集概述及其连接处
个人认为这篇文章内容十分丰富，并且非常适合初学者对视觉位置识别（Visual Place Recognition, VPR）这一领域中深度学习方法的理解。
文章详细介绍了多种成熟的方法，并从特征提取、匹配算法、多模态融合技术和多传感器集成等方面展开论述。
随后提供了详细的表格来对比各种方法及其对应的基准数据集。
文章不仅列举了多个公开可用的数据集，并提供了每个数据集的独特特点及获取链接。
关于每种方法的具体性能指标，则会在相应的基准数据集上进行评估描述。
总体而言，这篇文章内容详实且具有较高的参考价值。
Abstract
Visual place recognition has garnered significant attention across various fields, including computer vision and robotics. More recently, researchers have turned to advanced deep learning techniques to address these challenges. While numerous studies have introduced a variety of innovative place recognition methods grounded in deep learning, there remains a paucity of comprehensive overviews that elucidate the extent to which deep learning has been utilized in this domain. This paper presents an extensive survey conducted by delving into over 200 references, offering a thorough exploration from the perspective of deep learning. The paper commences with an overview of fundamental concepts in deep learning and their potential for location recognition. Subsequently, it delves into existing approaches built upon convolutional neural networks, encompassing both off-the-shelf models and those specifically designed for this task, alongside contemporary image representation techniques. Furthermore, the paper addresses the multifaceted challenges inherent in place recognition and conducts an exhaustive review of relevant datasets. To chart future research directions, the paper identifies open issues and introduces emerging tools such as generative adversarial networks, semantic scene understanding, and multi-modal feature learning. Finally, the paper concludes with a synthesis of these insights.
文章结构
-
Introduction
-
An introductory overview of deep neural networks
-
Deep learning: A conceptual framework
-
Subsection 2.1: An examination of fundamental concepts in deep learning
a. Deep learning: Understanding the core principles
b. Deep convolutional neural networks (DCNNs): Architectural innovations and applications -
Visual place recognition: based on a deep learning framework
3.1. Visual place recognition process
3.2. Visual place recognition as well as image retrieval -
基于CNN的位置识别技术
4.1 采用了经过训练的CNN模型
4.2 强大的图像表示能力
4.3 采用经过微调优化的CNN模型或新型架构设计
4.4 相似度衡量标准
4.5 评估标准
4.6 运行性能评估指标 -
Place recognition sets and complex challenges
5.1. Generic and prolonged sets
5.2. Specific sets and complex challenges -
Introducing new tools and addressing open challenges
6.1 Beyond the traditional convolutional neural networks
6.2 Semantic information analysis
6.3 Handling heterogeneous data -
Conclusion and research directions
几个有用的图表
- 视觉位置识别方法的时间发展线
- 视觉位置识别流程图
- CNN-based 方法比较
- 描述符比较（分全局和局部两种）
- 公开数据集一览表（原文中有附各数据集的地址）
