python comput vision framework construct imag retriev system focus user knowledg interact learn direct local pairwis base spars code effici method face retriev larg video dataset build compact local pairwis codebook joint featur space cluster build compact local pairwis codebook joint featur space cluster video retriev base track featur quantiz unifi approach detect identif commerci film tempor occurr pattern pagerank text similar video dduplic constraint new stori drank real time tunnel base video summar direct shift collis detect high dlevel featur extract svm walk dbase graph kernel person new video recommend new topic track drank queri expans base dduplic detect retak detect lc sift algorithm editori call amp calendareditori effici concept detect fuse simpl visual featur editori call amp calendareditori advanc multimedia model abstract attent psycholog measur human reflect stimulu propos gener framework highlight detect compar attent intens watch sport video step involv adapt select salient featur unifi attent estim highlight identif adapt select comput featur correl decid optim set salient featur unifi estim combin featur techniqu multi resolut auto regress mar creat tempor curv attent intens rank intens attent discrimin boundari highlight framework allevi semant uncertainti sport highlight lead effici effect highlight detect advantag capabl data coars tempor resolut robust nois caus modal asynchron percept uncertainti featur mismatch employ markovian constrain content present multi resolut estim attent intens enabl precis alloc event boundari kei word highlight detect attent comput sport video analysi scene duplic detect base pattern discontinu featur point trajectori imag dbase quiz gener new video archiv base princip object unsupervis face annot mine web approach hierarch classifi train multi dlevel imag annot approach filter junk imag googl search result approach hierarch classifi train multi dlevel imag annot text segment base approach video shot boundari detect ism trecvid high dlevel featur extract studi method support vector machin svm walk base graph kernel high level featur extract hlf task method imag segment finit set homogen segment repres segment graph vertex segment edg connect adjac segment set featur segment obtain posit definit kernel imag compar walk respect segment graph imag classif carri svm base kernel submit run method combin valu kernel svm paramet nation institut informat japan trecvid abstract paper report experi trecvid task high level featur extract search contentbas copi detect high level featur extract task baselin featur color moment edg orient histogram local binari pattern svm classifi nearest neighbor classifi search task approach includ search baselin featur search concept suggest video copi detect task studi approach base pattern motion featur point trajectori match frame pair normal cross correl approach consid baselin approach evalu task high level featur extract method overview rush summar redund elimin approach media object user dcenter similar match mediawalk video archiv explor base time dseri semant structur video search multi dmodal cluster analysi visual dtextual mutual inform entropi inter dmodal document index imag dbase quizz new video archiv boost face retriev relev set correl cluster scene duplic detect video base trajectori featur point larg scale new video databas brows retriev inform visual nation institut informat mine larg dscale new video databas knowledg visual polytechn univers ent dboost boost entropi measur robust object detect larg dscale new video retriev visual topic concept semant video shot classif abstract automat semant classif video databas user search brows challeng research problem combin visual text modal kei issu bridg semant gap signal semant paper propos enhanc classif highlevel concept intermedi topic concept studi fusion strategi combin topic concept visual featur order outperform unimod classifi conduct experi trecvid collect show intermedi topic classifi bridg part semant gap detect high level concept face retriev broadcast new video fuse tempor intens inform robust object detect fast featur select huge featur set exploit topic thread structur new video archiv semi dautomat gener video summari identif detect scene base flash light pattern ent dboost boost entropi measur robust object detect analyz person inform new video hybrid classifi precis robust ey detect analyz person inform new video multi dstage approach fast face detect paper describ multi stage approach achiev fast robust face detect approach motiv work viola jone cascad classifi yield coars fine strategi significantli reduc detect time maintain high detect rate distinguish previou work fact stage ad quickli estim face candid region larger window size larger move step size ii propos svm classifi adaboost classifi stage studi effici reus haar wavelet featur select adaboost previou stage svm classifi combin adaboost svm classifi final system obtain fast robust detect face pattern quickli reject earlier layer small number promis face pattern robustli classifi layer extens experiment result demonstr propos system achiev promis result cook navi assist daili cook kitchen trackthem explor larg dscale new video archiv track human relat effici featur select method object detect abstract propos simpl effici featur select method base principl compon analysi pca svm base classifi idea select featur ax closest principl compon comput data distribut pca experiment result show propos method reduc dimension similar pca maintain origin measur mean decreas comput time significantli examin enhanc ring dstructur graphic search interfac base usabl test advanc multimedia inform process pcm advanc multimedia inform process pcm advanc multimedia inform process pcm topic thread structur larg dscale new video archiv gener histogram empir optim low dimension featur imag match kei imag extract new video archiv visual semant structur new video browser ident video segment detect mine larg dscale broadcast video archiv inter dvideo structur multimedia integr cook video index distinct dsensit nearest neighbor search effici similar retriev multimedia inform nearest neighbor nn search high dimension featur space wide similar retriev multimedia inform recent research result databas literatur reveal curiou problem high dimension space high dimension space high degre freedom point scatter distanc yield signific differ case nn indistinct point exist similar distanc make matter wors indistinct nn requir search cost search complet choos nn plenti strong candid order circumv harm effect indistinct nn paper present nn search algorithm determin distinct nn search oper enabl cut search cost distinguish distinct nn indistinct advantag benefici interact retriev system semant retriev larg dscale video databas imag text featur subject region segment dispar map imag retriev distinct dsensit nearest neighbor search effici similar retriev multimedia inform nearest neighbor nn search high dimension featur space wide similar retriev multimedia inform recent research result databas literatur reveal curiou problem high dimension space high dimension space high degre freedom point scatter distanc yield signific differ case nn indistinct point exist similar distanc make matter wors indistinct nn requir search cost search complet choos nn plenti strong candid order circumv harm effect indistinct nn paper present nn search algorithm determin distinct nn search oper enabl cut search cost distinguish distinct nn indistinct advantag benefici interact retriev system subject region segment dispar map imag retriev compar evalu face sequenc match content dbase video access paper present compar evalu match method face sequenc obtain actual video face inform import video new program drama movi accur face sequenc match enabl multimedia applic includ content base face retriev autom face annot video author face sequenc video subject variat light condit pose facial express difficulti face match order cope problem face sequenc match method propos extend face imag match tradit pattern recognit recent pattern recognit techniqu expect applic face sequenc extract actual video perform method evalu accuraci face sequenc annot method accuraci evalu consider amount actual drama video evalua actor factress annot system drama video face sequenc match certainti factor evalu util face inform enabl multimedia applic new pro gram drama movi order util face inform face match method plai import role face match difficult face inform variat facial expr sion pose light condit order cope problem introduc reject mecha nism face sequenc match expect face inform match except reject mechan take advan tage certainti factor evalu paper de scribe method robust face sequenc match reject mechan mediasi video search engin mevis applic multidimension index method massiv process multimedia inform actor factress annot system drama video actor factress identif drama video multimedia document support insid ahyd platform effici implement evalu robust face sequenc match dit name detect face new video ion futur worldwid trend broadcast featur close caption close caption text transcript new video addit emploi video caption detect recognit cnn headlin new primari sourc new experi imag sequenc transcript video caption inform sourc associ extract face extract candid correl time inform face similar inform video caption account supplementari inform associ face name integr advanc imag process natur languag process techniqu face sequenc extract similar evalu video extract transcript video caption recognit technolog aren highli accur integr result system achiev accur output video ocr index digit new librari recognit superimpos caption automat extract recognit new caption annot great locat topic interest digit new video librari achiev goal present techniqu call video ocr optic charact reader detect extract read text area digit video data paper address problem describ method video ocr oper suggest applic digit new archiv solv problem charact recognit video low resolut charact extrem complex background appli interpol filter multi frame integr charact extract filter charact segment perform recognit base segment method intermedi charact recognit result improv segment includ method locat text area text properti languag base postprocess techniqu increas word recogni sr dtree index structur nearest dneighbor search high ddimension point data dit associ face video paper propos approach extract meaning content inform video collabor integr imag understand natur languag process actual develop system associ face name video call new video knowledg sourc automat extract face associ content inform system infer unknown face imag guess face system paper explain method success match result reveal effect integr heterogen techniqu import real content inform extract video face associ materi base work support nation scienc foundat cooper agreement iri opinion find conclus recommend express materi dit name detect face video integr imag natur languag process develop system associ face name new video knowledg sourc system new video includ imag sequenc transcript obtain audio track close caption text system infer face output candid locat face new video accomplish task system extract face imag sequenc name transcript correspond kei person new topic propos system take full advantag advanc imag natur languag process imag process contribut extract face sequenc provid rich inform face associ process help select frontal view face face sequenc enhanc face identif requir process hand natur languag process effect sr dtree index structur high ddimension nearest neighbor queri recent similar queri featur vector wide perform content base retriev imag appli techniqu larg databas requir develop multidimension index structur support nearest neighbor queri cientli ss tree propos purpos outperform index structur tree tree import featur emploi bound sphere bound rectangl shape region demonstr paper bound sphere occupi larger volum bound rectangl high dimension data reduc search effici overcom drawback propos index structur call sr tree sphere rectangl tree integr bound sphere bound rectangl region sr tree intersect bound sphere bound rectangl incorpor bound rectangl permit neighborhood partit smaller region ss tree improv disjoint region enhanc perform nearest neighbor queri highdimension uniform data practic actual imag video similar index includ perform test result verifi advantag sr tree show sr tree outperform ss tree tree type video scene classif system base typic model databas autom gener electron librari base document imag understand rule learn method academ document imag process robust line draw understand incorpor effici close symbol extract paper present effici extract method close loop primit symbol line draw robust line draw derstand system incorpor extrac tion method graph search techniqu symbol extract method provid effici model predict verif fa ciliti quantit consider comput cost shown line draw ing understand system effici tract model fairli complic line draw ing preserv comput cost fairli small syntact approach databas construct method document imag understand rule gener support system draw understand interact user nrcessari artliev rfftcirnt powrrt draw irnderstandinpl orrlcr aliz lar scrip ptactir draw imag datalm resid draw untlrrsland syst llard hirild ran bottl ck walixin lipsp dat ahil es autornaf ic genpratitrt irxwrng ndrrstand system desir tvp prmpo draw understandir ly rn state transit model system perform understand process compli understand rule declar form handld rahili tnxrt inp prearnt draw nrl tandiilg system nerat syqtem man machin interact system storeq pt viott rpss user nnd avoid rrpeatink queri a comput learn ethod pappr alter introduct system pr sent embodi experiment systrm succ ssful pxp rirnrtit show reveal rttrctiv ss system descript abil draw imag understand framework state transit model strongli desir construct multimedia databas map engin draw databas intend effec tive origin draw imag recogn complet propos framework draw imag understand system achiev framework featur understand knowledg free understand mechan understand knowledg ob tain easili understand mechan commonli applic understand system base framework implement found kind draw appli framework easili easi appli paper line framework appli framework kind draw imag veal kind draw applic kind discuss extend framework appli draw pre sent 