计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (4): 791-805.DOI: 10.3778/j.issn.1673-9418.2111028

• 综述·探索 • 上一篇    下一篇

锚点机制在目标检测领域的发展综述

伏轩仪1, 张銮景2, 梁文科2, 毕方明1,+(), 房卫东3   

  1. 1.中国矿业大学 计算机科学与技术学院,江苏 徐州 2211162. 中国赛宝(山东)实验室,济南 250013
    2.中国科学院 上海微系统与信息技术研究所 无线传感网与通信重点实验室,上海 200050
  • 收稿日期:2021-11-04 修回日期:2022-01-05 出版日期:2022-04-01 发布日期:2022-01-17
  • 通讯作者: + E-mail: bifangming@126.com
  • 作者简介:伏轩仪(1996—),女,江苏泰兴人,硕士研究生,CCF学生会员,主要研究方向为计算机视觉、边缘计算。
    张銮景(1965—),男,山东济南人,主要研究方向为电子信息技术产品、信息化工程技术。
    梁文科(1967—),男,山东济南人,主要研究方向为电子信息技术产品、信息化工程技术。
    毕方明(1974—),男,江苏徐州人,博士,副教授,主要研究方向为智能信息处理、空间信息安全。
    房卫东(1971—),男,山东济南人,博士,教授,主要研究方向为物联网信息安全、无线传感网信息安全。

Review on Development of Anchor Mechanism in Object Detection

FU Xuanyi1, ZHANG Luanjing2, LIANG Wenke2, BI Fangming1,+(), FANG Weidong3   

  1. 1. College of Computer Science and Technology, China University of Mining and Technology, Xuzhou, Jiangsu 221116, China
    2. China Saibao (Shandong) Laboratory, Jinan 250013, China
  • Received:2021-11-04 Revised:2022-01-05 Online:2022-04-01 Published:2022-01-17
  • About author:FU Xuanyi, born in 1996, M.S. candidate, student member of CCF. Her research interests include computer vision and edge computing.
    ZHANG Luanjing, born in 1965. His research interests include electronic information technoltthogy products and information engineering te-chnology.
    LIANG Wenke, born in 1967. His research interests include electronic information technology products and information engineering technology.
    BI Fangming, born in 1974, Ph.D., associate professor. His research interests include intelligent information processing and spatial information security.
    FANG Weidong, born in 1971, Ph.D., professor. His research interests include Internet of things information security and information security in wireless sensor networks.

摘要:

目标检测是计算机视觉领域的基本任务。近年来,基于深度学习的目标检测研究发展十分迅速,锚点(anchor)机制广泛应用于主流目标检测器中。多尺度的锚点是检测器解决尺度问题的有效方法,但锚点策略也存在尺寸固定、模型鲁棒性差等问题。根据优化锚点设置和无锚点(anchor-free)两种不同思路在目标检测中的发展,进一步分类总结检测模型的优缺点。首先回顾anchor策略提出的背景及原理,介绍基于优化anchor设置的目标检测模型,总结anchor机制存在的问题,引出无锚点(anchor-free)系列模型。在基于关键点的anchor-free模型中,按照检测思路分为基于特定位置关键点的检测器和结合中心关键点回归预测的检测器,分类总结算法的优缺点和使用范围,结合COCO数据集上的检测指标进一步对比。最后在总结融合anchor-based和anchor-free的模型基础上探讨两类算法的本质区别,指出未来的研究方向。

关键词: 目标检测, 锚点, 关键点, 标签分配

Abstract:

Object detection is a basic task in the field of computer vision. In recent years, object detection based on deep learning has developed rapidly. Anchor-based mechanism is widely used in mainstream object detectors. Multi-scale anchor is an effective method for detector to solve scale problems, but anchor strategy has some problems such as fixed size and poor robustness of model. According to the development of two different ideas in object detection: optimized anchor point setting and anchor-free, advantages and disadvantages of the detection model are further classified and summarized. Firstly, the background and principle of anchor strategy are reviewed, the object detection model based on optimized anchor setting is introduced, the existing problems of anchor mechanism are summarized, and a series of anchor-free models are introduced. Anchor-free model based on key points can be divided into detectors based on key points at specific locations and detectors combined with regression prediction of central key points according to the detection idea, advantages and disadvantages of various algorithms and their application scope are summarized by classification, and the detection indicators on COCO dataset are further compared. Finally, the essential differences between two algorithms are discussed on the basis of summarizing and integrating anchor-based and anchor-free models, and the future research direction is pointed out.

Key words: object detection, anchor, key points, label assign

中图分类号: