Volume 89, Issue 12, December 2009, Pages 2354-2366

The worldiscoveredwithmillionsofcameraswitheachrecordingahugeamountof
video.Itisatime-consumingtasktowatchthesevideos,asmostofthemareoflittle
interest duetothelackofactivity.Videorepresentationisthusanimportanttechnology
to tacklewiththisissue.However,conventionalvideorepresentationmethodsmainly
focusonasinglevideo,aimingatreducingthespatiotemporalredundancyasmuchas
possible.Incontrast,thispaperdescribesanovelapproachtopresentthedynamicsof
multiplevideossimultaneously,aimingatalessintrusiveviewingexperience.Givena
mainvideoandmultiplesupplementaryvideos,theproposedapproachautomatically
constructsasynthesizedmulti-videosynopsisbyintegratingthesupplementaryvideos
into themostsuitablespatiotemporalportionswithinthismainvideo.Theproblemof
findingsuitableintegrationbetweenthemainvideoandsupplementaryvideosis
formulatedasthemaximumaposterior(MAP)problem,inwhichthedesiredproperties
related toalessintrusiveviewingexperience,i.e.,informativeness,consistency,visual
naturalness,andstability,aremaximized.Thisproblemissolvedbyusinganefficient
Viterbi beamsearchalgorithm.Furthermore,aninformativeblendingalgorithmthat
naturalizestheconnectingboundarybetweendifferentvideosisproposed.
The proposedmethodhasawidevarietyofapplicationssuchasvisualinformation
representation, surveillancevideobrowsing,videosummarization,andvideoadvertis-
ing. Theeffectivenessofmulti-videosynopsisisdemonstratedinextensiveexperiments
overdifferenttypesofvideoswithdifferentsynopsiscases.