Stepwise Metric Promotion for Unsupervised Video Person Re-Identification

Stepwise Metric Promotion for Unsupervised Video Person Re-Identification

ID:40384672

大小:1021.74 KB

页数:10页

时间:2019-08-01

Stepwise Metric Promotion for Unsupervised Video Person Re-Identification_第1页
Stepwise Metric Promotion for Unsupervised Video Person Re-Identification_第2页
Stepwise Metric Promotion for Unsupervised Video Person Re-Identification_第3页
Stepwise Metric Promotion for Unsupervised Video Person Re-Identification_第4页
Stepwise Metric Promotion for Unsupervised Video Person Re-Identification_第5页
资源描述:

《Stepwise Metric Promotion for Unsupervised Video Person Re-Identification》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、StepwiseMetricPromotionforUnsupervisedVideoPersonRe-identificationZimoLiu1DongWang2∗HuchuanLu3DalianUniversityofTechnology12∗3lzm920316@gmail.comwdice,lhchuan@dlut.edu.cnAbstractTheintensiveannotationcostandtherichbutunlabeledID001datacontainedinvideosmotivateus

2、toproposeanunsuper-visedvideo-basedpersonre-identification(re-ID)method.ID002Westartfromtwoassumptions:1)differentvideotrack-letstypicallycontaindifferentpersons,giventhatthetrack-letsaretakenatdistinctplacesorwithlongintervals;2)withineachtracklet,theframesarem

3、ostlyofthesameper-son.Basedontheseassumptions,thispaperproposeastep-wisemetricpromotionapproachtoestimatetheidentitiesoftrainingtracklets,whichiteratesbetweencross-cameratrackletassociationandfeaturelearning.Specifically,WeFigure1.AvideoframeontheMARSdataset[48]

4、.Wedrawuseeachtrainingtrackletasaquery,andperformretrievalinthisframetwotracklets(redandyellow)whichareactuallyinthecross-cameratrainingset.Ourmethodisbuiltonobservedatdistincttimestamps.Itisintuitivetoassumethatthereciprocalnearestneighborsearchandcaneliminate

5、thetwotrackletscanbetreatedasdifferentidentities.Thisobservationhardnegativelabelmatches,i.e.,thecross-cameranearestcontributestothemodelinitializationprocess.neighborsofthefalsematchesintheinitialranklist.Thetrackletthatpassesthereciprocalnearestneighborchecko

6、ver,sincevideoscontainmorepriorknowledgeabouttheisconsideredtohavethesameIDwiththequery.Exper-scenario,theinfluenceofbackgroundnoisecanbelargelyimentalresultsonthePRID2011,ILIDS-VID,andMARSweakened[42].datasetsshowthattheproposedmethodachievesverycom-Second,vide

7、otracklets,producedbypedestriandetec-petitivere-IDaccuracycomparedwithitssupervisedcoun-tionandtracking(Fig.1),arereliabledatasourceforunsu-terparts.pervisedlearningmethods.Thisprocessisfullyautomaticandunsupervised.Asimpliedintherecentsurvey[50],differenttrack

8、letscanbetreatedasdifferentidentities,as1.IntroductionlongasweassumethatthetrackletsarecapturedatdistinctPersonre-identification(re-ID),aimingtoretrieveaplacesorwithlonginter

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。