Insufficiencyoflabeledtrainingdataisamajorobstacleforautomaticallyannotatinglarge-scalevideodatabaseswithsemanticconcepts.Existingsemi-supervisedlearningalgorithmsbasedonparametricmodelstrytotacklethisissuebyincorporatingtheinformationinalargeamountofunlabeleddata.However,theyarebasedon