Rahmani_Learning_a_Non-Linear_2015_CVPR_paper

Rahmani_Learning_a_Non-Linear_2015_CVPR_paper

ID:40724157

大小:3.22 MB

页数:9页

时间:2019-08-06

Rahmani_Learning_a_Non-Linear_2015_CVPR_paper_第1页
Rahmani_Learning_a_Non-Linear_2015_CVPR_paper_第2页
Rahmani_Learning_a_Non-Linear_2015_CVPR_paper_第3页
Rahmani_Learning_a_Non-Linear_2015_CVPR_paper_第4页
Rahmani_Learning_a_Non-Linear_2015_CVPR_paper_第5页
资源描述:

《Rahmani_Learning_a_Non-Linear_2015_CVPR_paper》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、LearningaNon-linearKnowledgeTransferModelforCross-ViewActionRecognitionHosseinRahmani,andAjmalMianComputerScienceandSoftwareEngineering,TheUniversityofWesternAustraliahossein@csse.uwa.edu.au,ajmal.mian@uwa.edu.auAbstractThispaperconcernsactionrecognitionfromunseenandunknownv

2、iews.Weproposeunsupervisedlearningofanon-linearmodelthattransfersknowledgefrommul-tipleviewstoacanonicalview.TheproposedNon-linearKnowledgeTransferModel(NKTM)isadeepnetwork,withweightdecayandsparsityconstraints,whichfindsasharedhigh-levelvirtualpathfromvideoscapturedfromdiffe

3、r-Figure1:Existingcross-viewactionrecognitiontechniques[11,entunknownviewpointstothesamecanonicalview.The12,19,33,40]connectsourceandtargetviewswithasetoflinearstrengthofourtechniqueisthatwelearnasingleNKTMfortransformationsthatareunabletocapturethenon-linearmanifoldsallacti

4、onsandallcameraviewingdirections.Thus,NKTMonwhichrealactionslie.OurNKTMfindsasharedhigh-leveldoesnotrequireactionlabelsduringlearningandknowl-non-linearvirtualpaththatconnectsmultiplesourceandtargetedgeofthecameraviewpointsduringtrainingortesting.viewstothesamecanonicalview.N

5、KTMislearnedonceonlyfromdensetrajectoriesofsyn-theticpointsfittedtomocapdataandthenappliedtorealApracticalsystemshouldbeabletorecognizehumanvideodata.Trajectoriesarecodedwithageneralcode-actionsfromdifferentunknownandmoreimportantlyun-booklearnedfromthesamemocapdata.NKTMissca

6、lableseenviews.Oneapproachforrecognizingactionsacrosstonewactionclassesandtrainingdataasitdoesnotre-viewpointsistocollectdatafromallpossibleviewsandquirere-learning.ExperimentsontheIXMASandN-UCLAtrainaseparateclassifierforeachview.However,thisap-datasetsshowthatNKTMoutperform

7、sexistingstate-of-the-proachdoesnotscalewellasitrequiresalargenumberofartmethodsforcross-viewactionrecognition.labeledsamplesforeachview.Toovercomethisproblem,sometechniquesinfer3Dscenestructureandusegeometrictransformationstoachieveviewinvariance[4,8,23,29,39].Thesemethodso

8、ftenrequirerobustjointestimationwhich1.Introductionisstillanopenprobleminre

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。
相关文章
更多
相关标签