文档介绍
UnsupervisedLearningofLong-TermMotionDynamicsforVideosZelunLuo,BoyaPeng,De-AnHuang,AlexandreAlahi,LiFei-FeiStanfordUniversity{zelunluo,boya,dahuang,alahi,feifeili}@AbstractWepresentanunsupervisedrepresentationlearningap-proachthatcompactlyencodesthemotiondependenciesinvideos.Givenapairofimagesfromavideoclip,ourframeworklearnstopredictthelong-term3Dmotions.Toreducethecomplexityofthelearningframework,wepro-posetodescribethemotionasasequenceofatomic3DflowscomputedwithRGB-Dmodality.WeuseaRecur-rentNeuralNetworkbasedEncoder-Decoderframeworktopredictthesesequencesofflows.Wearguethatinor-derforthedecodertoreconstructthesesequences,theen-codermustlearnarobustvideorepresentationthatcaptureslong-termmotiondependenciesandspatial-temporalrela-tions.Wedemonstratetheeffectivenessofourlearnedtem-Figure1.We