2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记

Deep Learning for Person Re-identification: A Survey and Outlook

本文是夜盲团队在今年发表的综述类文章:

论文地址

https://arxiv.org/abs/2001.04193v1

AGW开源地址

https://github.com/mangye16/ReID-Survey
当前的问题及概述
作者调查了245篇近两三年的行人重识别(Person Re-identification)论文,分类为封闭世界ReID与开放世界ReID,综述了该方向的技术进展,对未来ReID技术发展给出了几个有价值的方向。下图可见,在是否是异质数据、标注是否完备、是否含有噪声等方面,开放世界ReID更接近实际应用。
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
本文做了3方面工作:
1)对现有的深度学习方法进行了深入全面的分析,讨论了它们的优势和局限性。
2)为未来的开发设计了一个新的强大的AGW基线和一个新的评估度量(mINP)。AGW实现了最先进的性能,在单一和交叉模态Re-ID任务。mINP为现有的CMC/mAP提供了一个补充指标。
3)讨论几个重要的研究课题,以缩小封闭世界和开放世界应用之间的差距,向现实世界的Re-ID系统设计迈出一步。
ReID综述
1.ReID技术的五大步骤:
1)数据收集;2)包围框生成;3)训练数据标注;4)模型训练;5)行人检索
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
2.CLOSED-WORLD PERSON RE-IDENTIFICATION
1)特征表示学习方法:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
a)全局特征,学习每个人图像的全局表示;
b)局部特征,学习部分聚合的局部特征;
c)辅助特征,利用辅助信息学习特征表示(分割,viewpoint,domain,GAN)
d)视频特征,利用多图像帧和时间信息学习视频表示.
2)度量学习中的loss函数设计:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
常用的3中loss:(a) Identity Loss (b) Verification Loss © Triplet Loss以及他们的组合
Id loss:常用交叉熵loss
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
Verification Loss:一种为contrastive loss,一种为binary verification loss
其中contrastive loss:其中,ij为正样本时,δij = 1,相反为0
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
带有交叉熵的binary verification loss:其中, fij = (fj − fj)2 表示i,j两样本的差分特征
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
Triplet loss:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
Online Instance Matching (OIM) loss:在线学习loss,一个内存库{vk, k = 1,2,··,c}包含存储的实例特性,其中c表示类号:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
3)重排序优化:由易到难进行re-ranking排序
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
4)closed-world setting dataset:
约11个dataset:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
评价指标为CMC(rank-n)或mAP
State-of-the-arts(SOTA) :
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记

3 OPEN-WORLD PERSON RE-IDENTIFICATION:
1)异质数据ReID
基于深度ReID;
文本到图像ReID;
可见光到红外ReID;
跨分辨率ReID;
2)端到端ReID
纯图像/视频的ReID;
多摄像头跟踪的ReID;
3)半监督和无监督的ReID
其中无监督ReID SOTA方法统计:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
4)噪声鲁棒ReID
5)开放集合ReID
4.AN OUTLOOK: RE-ID IN NEXT ERA
1)mINP:本文提出的新的评价标准
设计了一个计算效率的度量,即负惩罚(NP),它度量惩罚来找到最难的正确匹配:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
Rhard表示最难匹配样本的rank位置,Gi表示i的正确匹配总数。
为了与CMC和mAP的一致性,我们倾向于使用逆负惩罚(INP),一种NP的逆运算:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
在不同模型比较:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
2)提出了一个新的baseling:AGW
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
I:Non-local Attention (Att) Block
采用强大的非局部注意块得到所有位置特征的加权和:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
II:Generalized-mean (GeM) Pooling
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
III:Weighted Regularization Triplet (WRT) loss
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
跨模态ReID比较:
2020Arxiv之ReID:Deep Learning for Person Re-identification: A Survey and Outlook论文笔记
5.Under-Investigated Open Issues
1)Scalable Re-ID
实现快速检索、轻量级网络、根据硬件配置自适应地调整模型
2)Domain Generalized Re-ID
多域(摄像机)数据集、多质(多模态等)数据集、
3)Minimizing Human Annotation
少标注、学习虚拟数据
4)Dynamic Camera Network
5)Domain-Specific Architecture Design