资源描述:
《Pose-Driven Deep Convolutional Model for Person Re-Identification英文文献资料》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、Pose-drivenDeepConvolutionalModelforPersonRe-identificationChiSu1∗†,JianingLi1∗,ShiliangZhang1,JunliangXing2,WenGao1,QiTian31SchoolofElectronicsEngineeringandComputerScience,PekingUniversity,Beijing100871,China2NationalLaboratoryofPatternRecognition,InstituteofAutomation,ChineseAcademyofScie
2、nces,Beijing100190,China3DepartmentofComputerScience,UniversityofTexasatSanAntonio,SanAntonio,TX78249-1604,USAsuchi@kingsoft.com,kaneiri1024@gmail.com,{slzhang.jdl,wgao}@pku.edu.cn,jlxing@nlpr.ia.ac.cn,qi.tian@utsa.eduAbstract123647Featureextractionandmatchingaretwocrucialcompo-58912nentsin
3、personRe-Identification(ReID).Thelargeposede-1310formationsandthecomplexviewvariationsexhibitedbythe1411capturedpersonimagessignificantlyincreasethedifficulty126oflearningandmatchingofthefeaturesfrompersonim-374ages.Toovercomethesedifficulties,inthisworkwepropose89125aPose-drivenDeepConvolution
4、al(PDC)modeltolearn1310improvedfeatureextractionandmatchingmodelsfromend1411toend.Ourdeeparchitectureexplicitlyleveragesthehu-(a)(b)(c)(d)(e)(f)manpartcuestoalleviatetheposevariationsandlearnFigure1.Illustrationofpartextractionandposenormalizationrobustfeaturerepresentationsfromboththegloba
5、limageinourFeatureEmbeddingsub-Net(FEN).Responsemapsof14anddifferentlocalparts.Tomatchthefeaturesfromglob-bodyjoints(b)arefirstgeneratedfromtheoriginalimagein(a).alhumanbodyandlocalbodyparts,aposedrivenfeature14bodyjointsin(c)and6bodypartsin(d)canhencebeinferred.Thepartregionsarefirstlyrotate
6、dandresizedin(e),thennormal-weightingsub-networkisfurtherdesignedtolearnadaptiveizedbyPoseTransformNetworkin(f).featurefusions.Extensiveexperimentalanalysesandresult-sonthreepopulardatasetsdemonstratesignificantperfor-lengingsituationslikecomplexviewvariationsandlargemanceimprovementsofourmo
7、deloverallpublishedstate-posedeformationsonthecapturedpersonimages.Mostofof-the-artmethods.traditionalworkstrytoaddressthesechallengeswiththefollowingtwoapproaches:(1)representingthevisualap-pearanceofapersonusingcustomizedlocalinvariantfea-1.Introductio