hybrid activ learn cross ddomain video concept detect acm intern workshop dlarg dscale multimedia corpu explicit implicit concept dbase video retriev bipartit graph propag model explicit implicit concept dbase video retriev bipartit graph propag model compar evalu protocol kth dataset control tv gestur vision base user interfac enabl natur interact modal gestur interfac requir computation intens video process low latenc demonstr applic recogn gestur control tv oper accur recognit achiev descriptor call mosift explicitli encod optic flow appear featur mosift computation expens sequenti implement run time slower real time reduc latenc suffici interact applic implement runtim system exploit parallel inher video understand applic categori subject descriptor comput system organ special purpos applic base system softwar softwar engin multi dprong approach improv semant extract new video acm sigmm workshop web dscale multimedia corpu wsmc acm sigmm workshop web dscale multimedia corpu wsmc identifi new video ideolog perspect emphat pattern visual concept video content analysi princip goal propos design automat video process architectur enabl design develop effici multimedia applic summar index retriev research project area reli mpeg high level semant retriev propos architectur compat mpeg propos architectur annot mpeg standard emphas user relev feedback high level semant retriev task propos research built substanti bodi preliminari work serv proof concept propos architectur techniqu preliminari work includ develop automat cluster techniqu delaunai triangul automat video summari gener automat shot boundari detect address problem automat video index organ video inform retriev user interact multipl level work focus import area tempor video segment construct semant multimedia abstract multimedia constitut interest field applic semant web semant web reason access manag multimedia content context depend strongli semant descript time multimedia resourc constitut complex object descript involv requir foundat sound model practic order repres find low high level multimedia analysi make access semant web queri resourc tutori aim provid red thread issu give outlin semant web model reason contribut area semant multimedia fruit interact field comput scienc semant multimedia multimedia object ubiquit found web search googl yahoo imag dedic site flickr youtub orinth effici search informedia video retriev system reliabl video concept detect great effort made improv video concept detect continu progress report current evalu method confin carefulli annot domain forgiv reliabl state art concept classifi remain question adopt rigor evalu approach find concept classifi built mainstream approach unreli gener poorli domain train domain evid show svm base concept classifi learn memor posit train data behav close memori base model knn compar perform model examin properti reliabl concept classifi find classifi frequent concept bloat classifi classifi capabl learn pattern data tend reliabl paper contribut understand concept detect suggest heurist identifi reliabl concept classifi discuss solut improv concept detect reliabl joint topic perspect model ideolog discours abstract polar discuss polit social issu common mass user gener media comput base understand ideolog discours consid difficult undertak paper propos statist model ideolog discours ideolog set gener belief social share group peopl democrat republican major polit ideolog unit state propos model captur lexic variat due ideolog text topic due author speaker ideolog perspect cope conjugaci logist normal prior deriv variat infer algorithm model evalu propos model synthet data written spoken polit discours experiment result strongli support ideolog perspect reflect lexic variat new video portrai new event ideolog perspect vox populi annot measur intens ideolog perspect aggreg group judgment framework classifi adapt applic concept detect adapt supervis classifi semant concept detector domain data paper describ gener framework function level classifi adapt base regular loss minim directli modifi decis function exist classifi type classifi domain base limit label data domain data make effici flexibl framework extend framework adapt multipl classifi classifi weight exist classifi learn automat reflect util elabor concret adapt algorithm deriv framework adapt svm multi adapt svm toon adapt experi adapt semant concept detector video channel type adapt approach proven superior origin unadapt classifi build term accuraci label effort informedia trecvid explor frontier explor util fast dforward surrog bbc rush paper discuss detail approach produc video summari submit trecvid bbc rush summar task includ baselin method empir work produc trecvid rush summar task gave strong evid simpl method sampl frame excel coverag text inclus perform submiss trecvid investig effect junk frame remov includ comprehens audio track emphas pan zoom backfil reclaim space remov nois shot origin set result show base method provid excel coverag expect limit effect strategi improv user satisfact discuss provid insight futur video summari develop evalu work measur novelti redund multipl modal cross dlingual broadcast new videolymp real dtime evalu multimedia retriev system interact prototyp convinc audienc multimedia technolog impact dynam audiovisu natur multimedia applic demonstr commun appli scienc effect static descript journal public ideal multimedia demonstr grasp audienc attent present effect result advanc multimod interfac mean user interact combin thereof interact demo offer research mean engag audienc achiev written manuscript major multimedia confer adopt demo session research equip laptop instal system show demo confer attende basi effect demo session lack common denomin individu system provid solut task data set confer attende familiar peculiar individu demo demo session make infeas present system entir confer audienc addit due lack focu impact demo session audienc suboptim editor note video search experi sens result tradit inform retriev metric fulli measur qualiti video search system provid interact assess todai video search engin author organ videolymp real time evalu showcas system compet answer specif video search front live audienc videolymp hear believ john smith establish focu good establish focu let demo system solv problem similar data set notabl focus effort take place demo session nation institut standard multimod new stori cluster pairwis visual dduplic constraint novelti detect cross dlingual new stori visual duplic speech transcript overwhelm volum new video channel languag todai demand automat manag abund inform effect search retriev brows track cross lingual new stori new stori similar measur plai critic role assess novelti redund paper explor novelti redund detect visual duplic speech transcript cross lingual new stori new stori repres sequenc keyfram visual track set word extract speech transcript audio track major differ pure text document number keyfram stori small compar number word exist larg number duplic keyfram featur make behavior similar measur compar tradit textual collect textual featur visual featur complement new stori combin boost perform experi trecvid cross lingual new video corpu show approach textual featur visual featur demonstr perform measur visual featur effect cosin distanc keyfram robust measur languag model built visual featur demonstr promis perform fusion textual visual featur improv perform cross ddomain video concept detect adapt svm practic elimin dduplic web video search queri expans probabilist local feedback applic multimedia retriev high dlevel concept fill semant gap new video retriev video retriev adapt svm classifi data shift distribut data mine applic benefit adapt exist classifi data shift distribut paper present adapt support vector machin adapt svm effici model adapt svm classifi train dataset dataset limit label exampl introduc regular svm object function adapt svm aim minim classif error train exampl discrep adapt origin classifi propos select sampl strategi base loss minim principl seed inform exampl classifi adapt experi artifici classif task benchmark video classif task show adapt svm outperform baselin method term accuraci effici adapt svm classifi data shift distribut data mine applic benefit adapt exist classifi data shift distribut paper present adapt support vector machin adapt svm effici model adapt svm classifi train dataset dataset limit label exampl introduc regular svm object function adapt svm aim minim classif error train exampl discrep adapt origin classifi propos select sampl strategi base loss minim principl seed inform exampl classifi adapt experi artifici classif task benchmark video classif task show adapt svm outperform baselin method term accuraci effici explor concept select strategi interact video search rank shot list autom lscom lite concept classifi investig respect trecvid topic select fit concept pair concept produc shot set greatest util draw fewer shot larger set concept averag precis measur show concept base shot set great util topic perfectli travers user empir data show realist abil separ relev shot irrelev recal relev topic depend perfect concept base strategi includ user driven select strategi ideal oracl priorit discuss implic queri concept interact video retriev concept space grow ten thousand hybrid approach improv semant extract new video paper describ hybrid approach improv semant extract new video experi show care paramet tune exploit multipl featur set multilingu linguist resourc appli text retriev approach imag featur establish synergi multipl concept undirect graphic model singl approach consist result concept detect suggest extract video semant exploit multipl resourc techniqu singl approach evalu bag dof dvisual dword represent scene classif discrimin field model semant concept video current think larg number semant concept provid research character video util video retriev understand semant concept isol exploit relationship multipl semant concept video sourc enhanc concept detect perform paper present discrimin learn framework call multi concept discrimin random field mdrf build probabilist model video semant concept detect incorpor relat concept observ propos model exploit power discrimin graphic model simultan captur associ concept observ data interact relat concept compar previou method model captur occurr concept incorpor data observ unifi framework present approxim paramet estim algorithm appli trecvid data experi show promis result compar singl concept learn approach video semant detect harmonium model semant video represent classif accur effici video classif demand fusion multimod inform intermedi represent combin idea framework propos probabilist approach video classif intermedi semant represent deriv multi modal featur base class bipartit undirect graphic model name harmonium approach repres video data latent semant topic deriv jointli model transcript keyword color histogram featur perform classif latent topic unifi framework show satisfactori classif perform approach benchmark dataset interest insight data provid approach summar bbc rush informedia clever cluster simpl speed dup summar rush high dlevel concept fill semant gap video retriev case studi broadcast new review text imag retriev approach broadcast new video document written perspect test perspect base statist distribut diverg paper investig automat determin document collect written perspect perspect point view perspect democrat republican propos test perspect base distribut diverg statist model collect experiment result show test successfulli distinguish document collect perspect type collect extrem video retriev joint maxim human comput perform present effici system video search maxim human bandwidth time exploit machin abil learn real time user select relev video clip system exploit human capabl rapidli scan imageri augment activ learn loop attempt present relev materi base current inform version human interfac evalu variabl page size manual page fix page size automat page requir absolut attent focu user optim perform case human search find relev result system invis rank previou guess number knowledg sourc imag similar text similar tempor proxim experiment evid show signific improv combin extrem human machin power approach wnew new video effici margin dbase rank learn algorithm inform retriev abstract learn good rank function plai kei role applic includ task multimedia inform retriev rank learn method explicitli model relat pair relev irrelev document result expens train process larg collect goal paper propos gener rank learn framework base margin base risk minim principl develop set effici rank learn approach model rank relat train time flexibl number margin base classifi extend rank learn counterpart rank logist regress develop paper experiment result show effici learn algorithm successfulli learn highli effect retriev function multimedia retriev trecvid collect effici margin dbase rank learn algorithm inform retriev abstract learn good rank function plai kei role applic includ task multimedia inform retriev rank learn method explicitli model relat pair relev irrelev document result expens train process larg collect goal paper propos gener rank learn framework base margin base risk minim principl develop set effici rank learn approach model rank relat train time flexibl number margin base classifi extend rank learn counterpart rank logist regress develop paper experiment result show effici learn algorithm successfulli learn highli effect retriev function multimedia retriev trecvid collect annot new video locat abstract locat video scene import semant descriptor broadcast new video paper propos learn base approach annot shot new video locat extract video transcript base featur multipl video modal includ syntact structur transcript sentenc speaker ident tempor video structur machin learn algorithm adopt combin multi modal featur solv problem locat video shot mention transcript locat transcript correct shot experi trecvid dataset demonstr approach achiev approxim accuraci correctli label locat shot new video explor synergi human machin extrem video retriev label disambigu sequenc model identifi human activ wearabl physiolog sensor mine relationship video concept probabilist graphic model larg scale automat semant video character learn model larg number semant concept semant concept exist isol exploit relationship multipl video concept sourc improv concept detect accuraci paper describ multi concept relat learn approach unifi probabilist graphic model represent propos numer graphic model mine relationship video concept appli perform video semant concept detect evalu compar trecvid video collect thousand word worth pictur experi video retriev thousand concept explor tempor consist video analysi retriev tempor consist ubiquit video data tempor adjac video shot share similar visual semant content paper present studi tempor consist defin respect semant concept queri topic quantit measur discuss implic video analysi retriev task show interact set tempor consist lead consider improv perform semant concept detect retriev video data specif activ learn method tempor sampl strategi propos build classifi semant concept tempor rerank method propos improv effici interact video search method outperform exist method consider margin trecvid dataset divers multimedia inform retriev research multimedia inform retriev highli divers field varieti data type research problem methodolog involv research field disciplin rang mathemat physic scienc comput scienc engin applic domain panel consist highli visibl activ research academia industri open discuss import divers healthi growth field paper record opinion express panel discrimin learn framework pairwis constraint video object classif video object classif insuffici label data time easili augment pairwis constraint sampl point class paper propos discrimin learn approach incorpor pairwis constraint convent margin base learn framework propos approach offer advantag exist approach deal pairwis constraint oppos learn distanc metric approach deriv classif power directli model decis boundari previou work handl label data convert pairwis constraint lead comput propos approach handl pairwis constraint label data comput greatli reduc propos approach evalu peopl classif task surveil video dataset probabilist latent queri analysi combin multipl retriev sourc combin output multipl retriev sourc document collect great import number retriev task multimedia retriev web retriev meta search merg retriev sourc adapt queri topic propos seri approach call probabilist latent queri analysi plqa associ ident combin weight latent class underli queri space compar previou queri independ queri class base combin method propos approach advantag discov latent queri class automat prior human knowledg assign queri mixtur queri class determin number queri class model select principl experiment result retriev task multimedia retriev meta search demonstr propos method uncov latent class train data achiev consider perform gain larg dscale concept ontolog multimedia dto sponsor lscom workshop develop larg concept lexicon multimedia includ approxim concept relat event object locat peopl program human subject annot concept corpu hour trecvid video compos subshot visual inspect keyfram singl imag subshot make judgment concept present absent keyfram keyfram approach fast proven suffici static concept airplan car person event activ concept airplan take car drive person walk tempor compon difficult human judg singl keyfram acquir label base view video motion make reliabl annot event total concept annot keyfram base method select annot video base approach videobas annot concept freeli download keyfram base label concept gather video base judgment design simpl tool provid binari posit neg label base view video clip subshot tool web base remot discrimin learn framework pairwis constraint video object classif video object classif insuffici label data time easili augment pairwis constraint sampl point class paper propos discrimin learn approach incorpor pairwis constraint convent margin base learn framework propos approach offer advantag exist approach deal pairwis constraint oppos learn distanc metric approach deriv classif power directli model decis boundari previou work handl label data convert pairwis constraint lead comput propos approach handl pairwis constraint label data comput greatli reduc propos approach evalu peopl classif task surveil video dataset learn rich semant new video archiv style analysi propos gener robust framework new video index found broadcast new product model identifi model product phase provid metadata annot contrast semiautomat index approach exploit inform product time adher automat data driven approach end analyz digit new video separ set multimod detector product phase combin result product deriv featur statist classifi ensembl framework facilit robust classif rich semant concept new video rich mean concept share similar product process experi archiv hour new video trecvid benchmark show combin analysi product phase yield result addit demonstr accuraci propos style analysi framework classif rich semant concept state art multipl instanc learn label face broadcast new video put activ learn multimedia applic dynam definit refin concept classifi author develop extens system video exploit put user control accommod situat sourc materi visual dens displai thumbnail imageri storyboard view shot base video explor retriev user identifi class audiovisu detect adeptli fluentli suppli train materi class iter evalu improv result automat classif produc multipl modal activ learn svm iter review output classifi updat posit neg train sampl effort typic relev feedback system user plai activ role direct classif process need truth small percentag multimedia data set exampl illustr iter creation classifi concept interest includ subsequ investig concept typic deem irrelev weed follow queri filter brows tool make exist iter ad concept put user control multimedia brows retriev process lesson futur decad informedia video analysi research util high dlevel semant featur video retriev abstract paper investig applic high level semant featur video retriev benchmark data trecvid address contribut featur outdoor face anim retriev user correctli decid featur appli pool truth data evid topic benefit featur studi subject found peopl disagre relev featur topic includ disagr posit featur topic associ strongli support truth data subject concur judgment correct topicfeatur pair identifi signific conduct investig interact search submiss show pair topic perform improv user access ideal classifi featur benefit deriv gener featur appli gener topic pair case specif featur appli specif topic rank submit shot base featur show promis automat search run interact run person care rank shot assess effect video retriev abstract paper examin result year trecvid video retriev evalu encourag evid progress video retriev major disappoint confirm field video retriev infanc public blith attribut improv retriev task techniqu pai attent statist reliabl comparison conduct analysi offici trecvid evalu result retriev experi error rate anova measur demonstr differ system statist signific conclud paper lesson learn result statist signific differ multi dmodal analysi person type classif new video classifi ident peopl appear broadcast new video anchor report new subject import topic high level video analysi visual resembl type peopl work explor multi modal featur deriv varieti evid speech ident transcript clue tempor video structur name entiti statist learn approach combin featur person type classif experi conduct abc world new tonight video demonstr effect approach contribut categori featur compar keyword multi modal analysi person type classif broadcast new video revisit effect topic set size retriev error multi dmodal analysi person type classif new video classifi ident peopl appear broadcast new video anchor report new subject import topic high level video analysi visual resembl type peopl work explor multi modal featur deriv varieti evid speech ident transcript clue tempor video structur name entiti statist learn approach combin featur person type classif experi conduct abc world new tonight video demonstr effect approach contribut categori featur compar keyword multi modal analysi person type classif broadcast new video mine text imag dual dwing harmonium propos multi wing harmonium model mine multimedia data extend improv earlier model base layer random field captur bidirect depend hidden topic aspect observ input model view undirect counterpart layer direct model lda similar task bear signific di erenc infer learn cost tradeoff latent topic represent topic mix mechan model facilit effici infer robust topic mix potenti high flexibl model latent topic space contrast diverg variat algorithm deriv learn special model dual wing harmonium caption imag incorpor multivari poisson word count multivari gaussian color histogram present empir result applic model classif retriev imag annot new video collect report extens comparison extant model learn queri dclass depend weight automat video retriev combin retriev result multipl modal plai crucial role video retriev system automat video retriev system user feedback queri expans current system util queri independ combin reli explicit user weight work propos queri class depend weight hierarchi mixtur expert framework combin multipl retriev result classifi user queri predefin categori aggreg retriev result queri class weight learn develop data effici gener unseen queri easili experiment result demonstr perform queri class depend weight consider surpass queri independ weight multimedia inform retriev workshop report proceed mdde learn queri dclass depend weight automat video retriev combin retriev result multipl modal plai crucial role video retriev system automat video retriev system user feedback queri expans current system util queri independ combin reli explicit user weight work propos queri class depend weight hierarchi mixtur expert framework combin multipl retriev result classifi user queri predefin categori aggreg retriev result queri class weight learn develop data effici gener unseen queri easili experiment result demonstr perform queri class depend weight consider surpass queri independ weight name individu new video monologu name individu person appear broadcast new video name detect video transcript lead access new video content paper approach challeng problem statist learn method categori inform extract multipl video modal explor featur distinguish true person constraint reveal relationship name person personnam problem formul learn framework predict person base featur refin predict constraint experi conduct abc world tonight cnn headlin new video demonstr approach outperform nonlearn altern larg amount success approach trec video retriev evalu paper review success approach evalu video retriev year task involv search retriev shot mpeg digit video record combin automat speech imag video analysi inform retriev technolog search evalu group interact human loop noninteract human enter queri system submiss interact search approach reli extens text retriev recent imag base featur contribut reliabl improv search perform interact approach substanti outperform interact approach system reli heavili user abil refin queri reject spuriou answer examin success automat search approach user interfac techniqu enabl high perform video retriev new associ new video word text retriev broadcast new video unsatisfactori transcript word frequent directli describ shot spoken extend retriev region window match keyword recal low precis improv text retriev approach segment visual stream coher stori unit set visual new stori delimit filter irrelev class shot left ambigu word transcript relat visual content remain shot stori limit set visual featur di erent semant level rang color histogram face car outdoor associ matrix captur correl visual featur specif transcript word matrix refin em approach preliminari result show approach potenti significantli improv retriev perform text queri find person correl name visual appear abstract peopl new subject carri rich semant broadcast new video find name person video major challeng video retriev task achiev exploit multi modal inform video includ transcript video structur visual featur propos comprehens approach find specif person broadcast new video explor clue name occur transcript face inform anchor scene importantli time pattern name peopl experi trecvid dataset show approach achiev high perform dretriev boost rerank approach video retriev larg scale concept ontolog broadcast video discrimin learn framework pairwis constraint video object classif video object classif insuffici label data time easili augment pairwis constraint sampl point class paper propos discrimin learn approach incorpor pairwis constraint convent margin base learn framework propos approach offer advantag exist approach deal pairwis constraint oppos learn distanc metric approach deriv classif power directli model decis boundari previou work handl label data convert pairwis constraint lead comput propos approach handl pairwis constraint label data comput greatli reduc propos approach evalu peopl classif task surveil video dataset detect tv new monologu style analysi propos method detect semant concept produc video base style analysi recognit concept appli classifi ensembl detect style element case studi present method detect concept new subject monologu approach averag precis perform submiss trecvid benchmark robust face recognit multipl view paper present approach aid face recognit multipl view face construct model directli imag recognit framework design video instanc target face sequenc slightli differ view oppos singl static pictur face specif reconstruct face shape orthogon view select featur base pairwis distanc landmark point model fisher linear discrimin face shape reconstruct sensit qualiti featur point local experi show reconstruct regular fisher linear discrimin provid highli accur face recognit multipl facial view experi carnegi mellon pie pose illumin express databas peopl face express vari light condit demonstr vastli improv perform merg rank list multipl sourc video classif multimedia corpora increasingli consist data multipl sourc di erent characterist exploit special applic paper focus video classif multipl sourc collect address question classifi train individu sourc full data set sourc train separ rank list di erent sourc merg ectiv formul problem merg rank list learn function map local score global score propos learn method base logist regress experi find sourc characterist import video classif method learn map function perform signific merg method explicitli learn map function model time featur broadcast new video classif broadcast new program structur video time strong predictor specif type new report learn classifi time featur easi task train data noisi paper approach problem gener model perspect approxim class densiti parametr fashion result show time simpl extrem ectiv featur method achiev significantli perform discrimin classifi multi dclass activ learn video semant featur extract activ learn demonstr tool reduc human label effort multimedia applic handl larg video collect previou work activ learn focus binari classif greatli limit applic activ learn paper present multi class activ learn approach extend activ learn binari classif multi class classif unifi represent margin base loss function experiment result trec semant featur extract task show propos activ learn approach work effect significantli reduc amount label data dine activ analysi hidden markov model describ algorithm dine activ analysi nurs home base featur includ motion vector distanc move region subspac individu person hidden markov model propos character stage dine activ tempor order hmm model identifi start end individu dine event high accuraci low fals posit rate approach success assist caregiv assess resid activ level time multi dmodal classif digit new librari video grammar locat name peopl find name person broadcast new video import video retriev reli text inform video transcript ocr text task suffer tempor mismatch person visual appear occurr text explor video grammar concurr pattern face name propos extend text base ir method overcom problem yield superior perform combin limit multimedia retriev combin search result multimedia sourc crucial deal heterogen multimedia data multimedia retriev final rank list item interest return sort confid relev attent combin function upper bound perform limit paper present theoret framework studi upper bound type combin function gener upper bound approxim propos monoton combin function studi upper bound linear combin function global optim techniqu experiment result show choic combin function consider influenc retriev perform neg pseudo drelev feedback content dbase video retriev video inform retriev requir system find inform relev queri repres simultan di erent wai text descript audio imag video sequenc present approach pseudo relev feedback retriev item similar queri item inquir user feedback provid insight approach statist model suggest score combin scheme posterior probabl estim evalu trec video track queri show techniqu improv video retriev perform real collect neg pseudo relev feedback show great promis difficult multimedia retriev task combin di erent retriev algorithm multimedia search pseudo drelev feedback present algorithm video retriev fuse decis multipl retriev agent text imag modal normal combin evid paper emphas success neg pseudo relev feedback improv imag retriev perform solv problem video inform retriev result encourag indic pseudo relev feedback show great promis multimedia retriev vari error data video retriev mix text imag queri paper present algorithm task video retriev queri consist text descript imag video pose video collect relev shot retriev system accomplish fuse retriev result multipl retriev agent system decompos agent includ text orient retriev agent respons find text speech transcript video ocr inform retriev ocr document content dbase probabilist correct model modifi logist regress approxim svm applic larg dscale text categor logist regress lr wide statist year receiv extens studi machin learn commun recent due close relat support vector machin svm adaboost paper modifi version lr approxim optim svm sequenc unconstrain optim problem prove approxim converg svm propos iter algorithm call mlr cg conjug gradient loop multiclass version mmlr cg obtain simpl modif compar mlr cg svmlight text categor collect show algorithm effici svmlight number train exampl larg result multiclass version mmlr cg report automat label video data multi dclass activ learn label video data essenti prerequisit vision applic depend train data visual inform retriev object recognit human activ model manual creat label time consum subject human error eventu imposs larg amount data surveil video minim human effort label propos unifi multi class activ learn approach automat label video data contribut paper includ extend activ learn binari class multipl class evalu practic sampl select strategi experiment result show propos approach work effect significantli reduc amount label data sampl select strategi achiev error reduct random sampl select faster iter scale algorithm condit exponenti model condit exponenti model wide condit model machin learn field improv iter scale ii major algorithm find optim paramet condit exponenti model paper propos faster iter algorithm name fi find optim paramet faster ii algorithm theoret analysi show propos algorithm yield tighter bound tradit ii algorithm empir studi text classif dataset show iter scale algorithm converg substanti faster ii algorithm conjug gradient algorithm cg examin qualiti optim paramet found learn algorithm case incomplet train experi shown limit amount comput allow converg achiev algorithm fi obtain lower test error ii method cg method modifi logist regress approxim svm applic larg dscale text categor logist regress lr wide statist year receiv extens studi machin learn commun recent due close relat support vector machin svm adaboost paper modifi version lr approxim optim svm sequenc unconstrain optim problem prove approxim converg svm propos iter algorithm call mlr cg conjug gradient loop multiclass version mmlr cg obtain simpl modif compar mlr cg svmlight text categor collect show algorithm effici svmlight number train exampl larg result multiclass version mmlr cg report constant densiti displai divers sampl problem remov subset imag visual complex object crowd visual order increas intellig occlus sever reduc imag intellig empti space convei imag inform optim choic tend constant densiti displai previou constant densiti algorithm base global densiti lead occlus empti space densiti uniform paper introduc algorithm consid layout individu object avoid occlus altogeth effici concern import dynam summari informedia digit video librari hundr thousand shot new stori post multipl queri account paramet visual origin queri reduc amount work requir video retriev speech imag inform video multipl type audio visual inform difficult extract combin trade gener video inform retriev paper evalu effect type inform video retriev video collect number sourc inform present typic broadcast video collect exploit inform retriev discuss contribut automat recogn speech transcript imag similar match face detect video ocr context experi perform part trec video retriev track evalu perform nation institut standard technolog queri evalu imag match video ocr prove decid aspect video inform retriev web imag retriev drank relev model web imag retriev challeng task requir effort imag process link structur analysi web text retriev content base imag retriev consid difficult current larg scale web imag search engin exploit text link structur understand content web imag local text inform caption filenam adjac text reliabl inform global inform account web imag retriev system make relev judgment paper propos rank method improv web imag retriev reorder imag retriev imag search engin rank process base relev model probabilist model evalu relev html document link imag assign probabl relev experi result show rank imag retriev achiev perform origin web imag retriev suggest effect rank method relev model learn internet prepar train data independ underli algorithm imag search engin rank process applic imag search engin effort new video classif svm dbase multimod classifi combin strategi video classif step multimedia content understand video classifi conceptu categori desir combin evid multipl modal combin strategi previou studi ad hoc investig meta classif combin strategi support vector machin compar probabl base strategi text featur closedcapt visual featur imag combin classifi broadcast new video experiment result show combin multimod classifi significantli improv recal precis meta classif strategi precis approach take product posterior probabl collag dynam summari new video paper introduc video collag effect interfac brows interpret video collect paper discuss collag automat produc illustr evalu effect summari new stori collag present text imag deriv multipl video sourc provid interact visual set video document summar content provid navig aid explor dynam creation collag base user context origin queri coupl automat process refin candid imageri name entiti identif common phrase extract descript text dynam manipul collag user direct brows reveal addit detail util collag summari examin respect publish new summari probabilist model titl gener titl gener complex task involv natur languag understand natur languag synthesi paper propos probabilist model titl gener previou statist model titl gener treat titl gener gener process convert document represent inform directli titl represent inform model introduc hidden state call inform sourc divid titl gener step step distil inform sourc observ document step gener titl estim inform sourc experi probabilist model outperform previou model titl gener term automat evalu human judgment trec video track inform retriev digit video inform paper intend provid comprehens pictur approach trec video track particip give overview trec video search task thumbnail sketch approach group reason write paper highlight messag trec video track varieti approach search brows digit video archiv approach work scalabl larger archiv yield retriev perform user import implic make digit librari video inform attain probabilist sourc model compar imag paper propos probabilist model imag retriev obtain similar queri imag imag collect model comput probabl gener imag observ queri imag compar probabilist model imag retriev color histogram base imag retriev method ibm qbic imag search engin evalu hour video retriev collect extract imag queri trec inform retriev evalu experiment result show probabilist model dramat outperform color histogram base imag retriev method ibm qbic imag search engin probabilist model camera zoom detect camera motion detect essenti autom video analysi propos probabilist model detect zoom zoom oper model em estim probabl zoom versu nonzoom oper standard mpeg motion vector tradit method set empir threshold deriv paramet proport zoom pan rotat tilt contrast probabilist model solid probabilist foundat clear simpl probabl threshold experi show probabilist model significantli perform baselin parametr method zoom detect precis recal video dcuebik adapt imag search video shot propos analysi search imag video librari simpl imag search compar imag frame kei idea expand definit imag account variabl sequenc video frame compris shot implement method qbic imag search engin show clear improv imag search combin tradit imag search video imag search provid result trec video retriev evalu data multi dmodal inform retriev broadcast video ocr speech recognit examin multi modal inform retriev broadcast video text read screen ocr speech recognit perform audio track ocr speech recognit compar trec video retriev evalu corpu result show ocr import speech recognit video retriev ocr retriev improv dictionari base post process demonstr util imperfect multi modal metadata result benefit multi modal inform retriev wearabl digit librari person convers develop wearabl person digit librari system unobtrus record wearer part convers recogn face current dialog partner rememb voic time system see person face hear voic replai audio convers compress form summar name major issu mention experi prototyp system show combin face recognit speaker identif effect retriev convers video retriev multipl imag search strategi meta dclassif multimedia classifi combin multipl classifi interest multimedia system data type modal mine analyz wearabl experi collect system unobtrus record wearer convers recogn face dialog partner rememb voic system see person face hear voic summari convers person remind wearer correctli identifi person mixtur video audio stream classif judgment individu modal classifi combin effect yield accur decis address problem combin strategi previou studi meta classif strategi support vector machin propos preliminari result show combin face recognit speaker identif technolog meta classif dramat effect weight interpol meta classif gener appli applic combin multipl classifi modif meta dclassif combin multimod classifi languag model ir collect inform earli ag ir document repres control vocabulari paper explor usag meta data inform framework languag model present languag model advantag categori inform document improv retriev accuraci compar languag model tradit languag model trec dataset collect inform document obtain kmean cluster method languag model outperform tradit languag model verifi statement titl languag model inform retriev paper propos languag model titl languag model inform retriev tradit languag model retriev defin condit probabl probabl queri titl document adopt statist translat model learn titl document pair collect comput probabl avoid spars data problem propos smooth method experi trec document collect titl languag model inform retriev smooth method outperform tradit languag model vector space model ir significantli video classif retriev informedia digit video librari system paper organ part part detail lower level shot classif work part describ manual retriev system section detail interact retriev system carnegi mellon univers trec video retriev track run descript data found proceed trec confer video track overview classif trec video track main task detect semant featur concept indoor outdoor peopl part descript classif task submit run classif concept trec video track obtain train data manual annot frame hour featur develop collect categori outdoor segment recogniz outdoor locat build exclud scene indoor close up object object outdoor indoor segment recogniz indoor locat insid build exclud scene outdoor close up object object indoor cityscap segment recogniz citi urban suburban set monologu event singl person partial visibl speak long time titl gener train corpu learn select good titl word approach base revers inform retriev paper show learn select good word document titl view problem select good titl word document variant inform retriev problem titl word treat document select titl word find relev document base train collect consist document titl pair learn document represent titl word appli learn represent select titl word test document compar learn approach nearest neighbor approach nave bayesian approach variant machin translat model find approach significantli metric titl gener machin dtranslat document paper present compar automat gener titl machin translat document statist base method na ve bayesian nearest neighbour tf idf iter expect maxim method titl gener appli origin english new document document translat english portugues french german back english systran autosummar function microsoft word base line result metric show statisticsbas method titl gener machinetransl document fairli languag independ titl gener level approach accuraci titl gener origin english document demonstr hierarch document cluster digit librari retriev result digit librari grow size queri content frustrat queri web remedi hierarch cluster result return search digit librari demonstr cluster search result carnegi mellon informedia databas larg video librari support index retriev automat gener descriptor meta dscore automat evalu term weight scheme ir precis drecal paper present method automat evalu perform term weight scheme inform retriev resort precis recal base human relev judgment specif problem document term matrix gener term weight scheme term weight scheme perform propos meta score function take input document term matrix gener term weight scheme comput good score documentterm matrix experi found score highli correl precis recal measur collect term weight schema conclud meta score function substitut precis recal measur relev judgment human subject meta score function limit text inform retriev appli field imag dna retriev video retriev informedia digit video librari system inform extract video audio content terabyt onlin data collect automat gener metadata indic retriev video librari architectur project base premis real time constraint librari metadata creation relax order realiz increas autom deeper pars index identifi librari content break segment librari creation offlin activ librari explor user occur onlin make gener metadata segment goal informedia interfac enabl quick access relev inform digit video librari leverag deriv metadata partit video small segment figur show idvl interfac queri figur set result displai bottom displai includ window headlin pictori menu video segment repres thumbnail imag approxim resolut video horizont vertic dimens headlin window automat pop mous posit result item headlin window result shown idvl support wai navig brows digit video librari interfac featur essenti deal ambigu deriv data gener speech recognit automat titl gener em data analysi multimedia librari complementari video audio analysi broadcast new archiv adjust filmstrip skim abstract digit video librari filmstrip video skim present scheme abstract inform digit video segment filmstrip present inform static form video skim plai disclos inform tempor paper discuss evolut filmstrip skim interfac informedia digit video librari filmstrip commonli deploi interfac video imag librari found initi informedia filmstrip skim receiv discuss interfac consider motiv redesign filmstrip skim adjust present dynam base user context prefer improv acoust model caption multimedia speech cmu spoken document retriev trec analysi role term frequenc tf paper attempt tf idf theoret interpret point reason multipli tf directli idf poor perform interpret word histogram base method integr tf idf factor log probabl document tf occurr word success tf idf ln tf idf fail compet ln ln tf idf failur due end effect pursu problem futur informedia experi don ddemand captur informedia experi demand system speech imag natur languag process combin gp inform captur integr commun person multimedia experi paper discuss initi prototyp eod system lesson learn build terabyt digit video librari boundari manual documentari video subsequ experi look gener stori boundari automat stori video segment averag minut length entir librari contain video segment artifici intellig techniqu creat metadata data describ video content found ai techniqu applic new corpu documentari corpu techniqu includ speech recognit imag process inform retriev learn integr area compens limit accuraci coverag commun power speech process analyz audio compon video cmu sphinx speech recognit system creat complet transcript text base retriev speech align exist imperfect tran learn recogn speech watch televis topic label broadcast new stori informedia digit video librari stori segment detect commerci broadcast new video informedia digit librari project wactlar full content index retriev text audio video materi segment integr process informedia digit video librari success informedia project hing critic assumpt extract suffici accur speech recognit transcript broadcast audio segment broadcast video paragraph stori inform retriev previou paper hauptmann witbrock witbrock shown speech recognit suffici inform retriev pre segment video new stori paper address issu segment demonstr fulli automat system extract stori boundari audio video close caption cue stori segment step informedia digit video librari split full length new broadcast individu new stori pha experi spoken document retriev cmu describ submiss trec spoken document retriev sdr track speech recognit inform retriev engin present sdr evalu result analysi develop greater detail includ probabilist retriev engin base languag model tfidf base weight function incorpor word error probabl simpl confid estim word probabl base speech recognit lattic improv develop test set promis techniqu fail yield signific gain evalu test set sdr data task entir set speech data trec spoken document retriev track consist hour broadcast new approxim train test data segment stori manual transcrib test set version data manual gener transcript speech recognit transcript base ibm cmu recogn raw audio data transcrib recogn entir train set train acoust model speech recognit system remaind held unseen test data stori train data set test set develop debug system trec evalu set item retriev system queri document assign relev experi evalu test set averag precis retriev relev document judg qualiti retriev relev judgement develop test se speech recognit digit video librari standard method make full content audio video materi searchabl annot human gener meta data describ content search understand creation multimedia cd rom huge amount data usefulli includ digit video audio librari cost produc meta data prohibit informedia digit video librari product meta data support librari interfac autom techniqu deriv artifici intellig ai research appli speech recognit natur languag process inform retriev imag analysi interfac produc help user locat inform navig brows digit video librari effect specif interfac compon includ automat titl filmstrip video skim word locat mark repres frame word phonet string effici inform retriev imperfectli transcrib spoken document artifici intellig techniqu interfac digit video librari experi spoken document retriev cmu describ submiss trec spoken document retriev sdr track speech recognit inform retriev engin present sdr evalu result analysi develop greater detail includ probabilist retriev engin base languag model tfidf base weight function incorpor word error probabl simpl confid estim word probabl base speech recognit lattic improv develop test set promis techniqu fail yield signific gain evalu test set sdr data task entir set speech data trec spoken document retriev track consist hour broadcast new approxim train test data segment stori manual transcrib test set version data manual gener transcript speech recognit transcript base ibm cmu recogn raw audio data transcrib recogn entir train set train acoust model speech recognit system remaind held unseen test data stori train data set test set develop debug system trec evalu set item retriev system queri document assign relev experi evalu test set averag precis retriev relev document judg qualiti retriev relev judgement develop test se demonstr read coach listen grant fee provid copi made distribut profit commerci advantag copi bear notic full citat page copyright compon work own acm honor abstract credit permit copi republish post server redistribut list requir prior specif permiss fee demonstr read coach listen speech multimedia inform retriev describ informedia tm new demand system new demand innov index search broadcast video audio materi text content fulli automat system monitor tv new select retriev new item base spoken queri user plai video paragraph system run pentium pc mpeg video compress sphinx ii continu speech recognit system read coach listen edit video transcript prototyp read coach listen grant fee provid copi made distribut profit commerci advantag copi bear notic full citat page copyright compon work own acm honor abstract credit permit copi republish post server redistribut list requir prior specif permiss fee demonstr read coach listen prototyp read coach listen summari project listen survei current speech technolog articl describ technolog speech recognit speech synthesi manipul speech term inform content recognit transform human speech text liter dictat interpret command control applic synthesi gener spoken utter text synthesi desir larg number utter messag content unpredict requir make pre record speech impract technolog cover articl interest support direct commun human comput mode human commonli commun highli skill speech technolog note discuss includ speaker recognit automat establish speaker ident speech edit index manipul speech extract linguist inform recognit read coach listen autom detect oral read error grant fee provid copi made distribut profit commerci advantag copi bear notic full citat page copyright compon work own acm honor abstract credit permit copi republish post server redistribut list requir prior specif permiss fee gestur speech graphic manipul experi conduct peopl gestur speech manipul graphic imag comput screen human substitut recognit devic analysi show peopl strongli prefer gestur speech graphic manipul intuit multipl hand multipl finger dimens surpris uniform simplic gestur speech analysi result strong encourag futur develop integr multi modal interact system keyword gestur speech manipul input graphic introduct paper present studi peopl intuit gestur speech commun comput explos capabl comput present output bandwidth input channel comput limit keyboard input devic user syntax mean natur languag process layer predict flexibl dialog expect speech recognit relax contextu constraint violat comput speech recognit problem order organ constraint set solv plan base task predict featur success restrict call layer user behavior infer aid constraint violat violat constraint recognit speech input mind system restrict layer reduc search space high level knowledg sourc usabl speech recognit system dialog dlevel knowledg sourc improv speech recognit motiv describ implement mind speech recognit system mind knowledg dialog structur user goal focu problem solv situat knowledg combin form predict translat dynam gener semant network grammar experi evalu recognit accuraci level knowledg constraint result show speech recognit accuraci improv dramat maxim constrain dynam network grammar process speech input signal talk comput empir investig pars spoken languag semant casefram approach pars spoken input introduc problem present pars type natur languag indeterminaci inaccuraci acoust recognit handl integr manner techniqu pars type natur languag adapt extra demand paper describ extens semant casefram pars restrict domain spoken input semant casefram grammar representalien earlier work robust pars type input due uncertainti inher speech recognit casefram grammar appli emphas island grow casefram header radic chang applic due high degrb abstract casefram represent eppro tch present test successfulli preliminari implement 