[ICML 2010] 3D Convolutional Neural Networks for Human Action Recognition

《[ICML 2010] 3D Convolutional Neural Networks for Human Action Recognition》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、3DConvolutionalNeuralNetworksforHumanActionRecognitionShuiwangJishuiwang.ji@asu.eduArizonaStateUniversity,Tempe,AZ85287,USAWeiXuxw@sv.nec-labs.comMingYangmyang@sv.nec-labs.comKaiYukyu@sv.nec-labs.comNECLaboratoriesAmerica,Inc.,Cupertino,CA95014,USAAbstractclutteredbackgrounds,occlusions,andviewpoi

2、ntvari-ations,etc.Therefore,mostoftheexistingapproachesWeconsiderthefullyautomatedrecognition(Efrosetal.,2003;Sch¨uldtetal.,2004;Doll´aretal.,ofactionsinuncontrolledenvironment.Most2005;Laptev&P´erez,2007;Jhuangetal.,2007)existingworkreliesondomainknowledgetomakecertainassumptions(e.g.,smallscalea

3、ndview-constructcomplexhandcraftedfeaturesfrompointchanges)aboutthecircumstancesunderwhichinputs.Inaddition,theenvironmentsarethevideowastaken.However,suchassumptionssel-usuallyassumedtobecontrolled.Convolu-domholdinreal-worldenvironment.Inaddition,mosttionalneuralnetworks(CNNs)areatypeofoftheseap

4、proachesfollowtheconventionalparadigmdeepmodelsthatcanactdirectlyontherawofpatternrecognition,whichconsistsoftwostepsininputs,thusautomatingtheprocessoffea-whichthefirststepcomputescomplexhandcraftedfea-tureconstruction.However,suchmodelsareturesfromrawvideoframesandthesecondsteplearnscurrentlylimi

5、tedtohandle2Dinputs.Inthisclassifiersbasedontheobtainedfeatures.Inreal-worldpaper,wedevelopanovel3DCNNmodelforscenarios,itisrarelyknownwhichfeaturesareimpor-actionrecognition.Thismodelextractsfea-tantforthetaskathand,sincethechoiceoffeatureisturesfrombothspatialandtemporaldimen-highlyproblem-depend

6、ent.Especiallyforhumanac-sionsbyperforming3Dconvolutions,therebytionrecognition,differentactionclassesmayappearcapturingthemotioninformationencodeddramaticallydifferentintermsoftheirappearancesinmultipleadjacentframes.Thedevelopedandmotionpatterns.modelgeneratesmultiplechannelsofinfor-mationfromthei

7、nputframes,andthefinalDeeplearningmodels(Fukushima,1980;LeCunetal.,featurerepresentationisobtainedbycom-1998;Hinton&Salakhutdinov,2006;Hintonetal.,bininginformationfromallchannels.We2006;Bengio,2009)areaclassofmac