基于关键对象选择摘要方法研究.pdfVIP

FromKeyframestoKeyObjects:

SummarizationbyRepresentativeObjectProposalSelection

{jingjing.meng,jsyuan,eyptan}@ntu.edu.sg,ihxwang@cqu.edu.cn

Weproposetosummarizeaintoafewkeyob-jects

byselectingrepresentativeobjectproposalsgeneratedfrom

frames.Thisrepresentativeselectionproblemis

formulatedasasparsedictionaryselectionproblem,.,

choosingafewrepresentativesobjectproposalstorecon-

structthewholeproposalpool.Comparedwithexisting

sparsedictionaryselectionbasedrepresentativeselection

methods,ournewformulationcanincorporateobjectpro-

posalpriorsandlocalitypriorinthefeaturespacewhen

selectingrepresentatives.Consequentlyitcanterlo-

catekeyobjectsandsuppressoutlierproposals.Weconvert

theoptimizationproblemintoaproximalgradientproblem

andsolveitbythefastiterativeshrinkagethresholdingal-

gorithm(FISTA).Experimentsonsyntheticdataandreal

benchmarkdatasetsshowpromisingresultsofourkeyob-

jectsummarizationapporachincontentminingand

search.Comparisonswithexistingrepresentativeselection

approachessuchasK-mediod,sparsedictionaryselection

anddensitybasedselectionvalidatethatourformulatione.g

cantercapturethekeyobjectsdespiteappearance

variations,clutteredbackgroundsandcameramotions.

1.Introductione.g