benchmark dynam time warp music retriev studi perform dynam program method music retriev method design time seri match directli appli retriev music dynam time warp dtw identifi optim align time seri comput match cost align signific speed up achiev constrain dynam time warp cdtw narrow set posit time seri match specif posit time seri method design full sequenc match appli subsequ match slide window databas sequenc comput match score databas subsequ addit spring dynam program approach design subsequ match queri match databas subsequ requir match length equal queri length spring lower comput cost dtw cdtw databas consist set midi file web midi file convert dimension time seri take account note pitch durat synthet queri fix size nois level surprisingli top match approach show similar behavior term retriev accuraci small valu suggest specif applic area computation cheaper method spring suffici retriev top match experi comput vision method fall detect databas dbase framework gestur recognit reduc jointboost dbase multiclass classif proxim search boost versu ova classifi commonli multiclass problem gener object recognit biometr base identif gestur recognit jointboost recent propos method ova classifi train jointli forc share featur jointboost demonstr lead higher accuraci smaller classif time compar ova classifi train independ share featur improv effici jointboost time complex ovabas multiclass recognit linear number class lead prohibit larg run time domain larg number class paper shown jointboost base recognit reduc classif time nearest neighbor search vector space reduct propos simpl easi implement vector index scheme base princip compon analysi pca experi propos method achiev speedup order magnitud standard jointboost classif hand pose recognit system number class close neglig loss classif accuraci method yield promis result experi wide frgc face recognit dataset number class similar measur vision dbase sign recognit abstract encount english word understand dictionari american sign languag asl user encount unknown sign mean sign straightforward process recent propos problem address comput vision system help user mean sign approach sign lookup treat video databas retriev problem user encount unknown sign user video sign queri retriev similar sign databas compon sign lookup system similar measur compar sign video queri video specif sign similar measur assign high similar valu video sign low similar valu video sign paper evalu state art video base similar measur call dynam space time warp dstw purpos sign retriev paper discuss specif adapt dstw toler differ translat scale autom larg vocabulari gestur search paper describ work design comput vision system help user mean sign sign lookup treat video databas retriev problem video databas util video exampl sign larg number sign close current experi emphasi paper evalu tradeoff autom approach user manual specifi hand locat input video fulli autom approach hand locat determin comput vision modul introduc inaccuraci sign retriev process experiment evalu approach present respect advantag disadvantag faster activ search embed dbase subsequ match event search problem identifi event activ interest larg databas store long sequenc activ paper topic problem identifi activ interest databas activ repres time seri typic setup user present queri repres activ interest system retriev similar activ store databas focu case databas match segment priori databas represent long continu activ occur extens period time queri constraint databas match start end longer activ pattern contain popular dtw measur databas match found dynam program retriev time linear size databas long databas size larger achiev effici retriev time appli problem recent propos techniqu call embeddingbas subsequ match ebsm demonstr ebsm obtain signific speedup retriev time unifi framework gestur recognit spatiotempor gestur segment abstract context hand gestur recognit spatiotempor gestur segment task determin video sequenc gestur hand locat gestur start end exist gestur recognit method typic assum spatial segment tempor segment paper introduc unifi framework simultan perform spatial segment tempor segment recognit propos framework inform flow bottom top gestur recogn hand locat highli ambigu inform gestur begin end unavail method appli continu imag stream gestur perform front move clutter background propos method consist contribut spatiotempor match algorithm accommod multipl candid hand detect frame classifi base prune framework enabl accur earli reject poor match gestur model subgestur reason algorithm learn gestur model fals match part longer gestur perform approach evalu challeng applic recognit hand sign digit gestur user wear short sleev shirt front clutter background retriev occurr sign interest video databas continu unseg sign american sign languag asl unifi framework gestur recognit spatiotempor gestur segment abstract context hand gestur recognit spatiotempor gestur segment task determin video sequenc gestur hand locat gestur start end exist gestur recognit method typic assum spatial segment tempor segment paper introduc unifi framework simultan perform spatial segment tempor segment recognit propos framework inform flow bottom top gestur recogn hand locat highli ambigu inform gestur begin end unavail method appli continu imag stream gestur perform front move clutter background propos method consist contribut spatiotempor match algorithm accommod multipl candid hand detect frame classifi base prune framework enabl accur earli reject poor match gestur model subgestur reason algorithm learn gestur model fals match part longer gestur perform approach evalu challeng applic recognit hand sign digit gestur user wear short sleev shirt front clutter background retriev occurr sign interest video databas continu unseg sign american sign languag asl refer dbase align larg sequenc databas paper introduc method call refer base string align rbsa speed retriev optim subsequ match larg databas sequenc edit distanc smith waterman similar measur rbsa oper assumpt optim match deviat small amount queri amount exce prespecifi fraction queri length rbsa exact version guarante fals dismiss handl larg queri effici approxim version rbsa achiev signific addit improv exact version neglig loss retriev accuraci rbsa perform filter candid match precomput align score databas sequenc set fix length refer sequenc queri time queri sequenc partit segment length equal refer sequenc segment align score segment refer sequenc effici identifi small number candid subsequ match alphabet collaps techniqu emploi improv prune power filter step experiment evalu rbsa significantli outperform state art biolog sequenc align method gram blast bwt track dynam hidden dstate shape model abstract hidden state shape model hssm previous propos repres detect object imag exhibit deform shape variat structur paper introduc dynam hidden state shape model dhssm track recogn rigid motion object human hand recurs bayesian filter method call dp track combin exhaust local search match imag featur model state dynam program approach find global registr model object imag contribut techniqu exploit hierarch structur dynam program approach averag consider speed search match propos emb onlin learn approach track mechan updat dhssm dynam learn approach ensur dhssm accur repres track object distinguish clutter potenti present imag experi show method recogn digit hand finger move curl degre method robust illumin condit presenc clutter occlus type occlus experi demonstr signific improv effici accuraci recognit compar recurs frame frame detect nearest neighbor retriev distanc dbase hash abstract method propos index space arbitrari distanc measur achiev effici approxim nearest neighbor retriev hash method local sensit hash lsh successfulli appli similar index vector space string space ham distanc kei novelti hash techniqu propos appli space arbitrari distanc measur includ metric distanc measur describ domain independ method construct famili binari hash function function construct multipl multibit hash tabl show lsh formal applic analyz behavior tabl index structur present formul statist observ sampl data analyz retriev accuraci effici propos index method experi real world data set demonstr method produc good trade off accuraci effici significantli outperform vp tree method distanc base index benchmark databas video dbase automat sign languag recognit linguist annot video databas automat sign languag recognit present rwth boston corpu consist sentenc speaker separ subset train develop test detail evalu benchmark automat sign languag recognit larg corpora need recent research focus isol sign languag recognit method video sequenc record lab condit special hardwar data glove databas consist gener speaker speaker depend small vocabulari databas access interfac design creat provid fast access databas statist content make easili brows retriev subset video databas preliminari baselin result corpora present contradistinct research area databas present paper publicli nearest neighbor search method handshap recognit gestur import modal human machin commun robust gestur recognit import compon intellig home assist environ gener import aspect gestur handshap handshap hold import inform mean gestur sign languag intent action manipul gestur virtual realiti interfac time recogn handshap challeng task handshap imag depend orient hand viewpoint camera paper examin databas approach handshap classif larg databas ten thousand imag repres wide variabl handshap appear effici accur index method import databas approach ensur system match incom imag larg number databas imag interact time paper examin embed base hash tabl base index method handshap recognit experiment compar approach task recogn handshap commonli american sign languag asl translat scale dinvari gestur recognit complex scene approxim embed dbase subsequ match time seri method approxim subsequ match introduc significantli improv effici subsequ match larg time seri data set dynam time warp dtw distanc measur method call ebsm shorthand embed base subsequ match kei idea convert subsequ match vector match embed embed map databas time seri sequenc vector step time seri databas map vector embed comput appli full dynam time warp refer object databas time seri runtim queri object embed object comput manner run dynam time warp refer object queri compar embed queri databas vector effici identifi area interest databas sequenc area interest fulli explor exact dtw base subsequ match algorithm experi larg public time seri data set produc speedup order magnitud compar brute forc search small loss lt retriev accuraci boostmap embed method effici nearest neighbor retriev abstract paper describ boostmap method effici nearest neighbor retriev computation expens distanc measur databas queri object embed vector space distanc measur effici embed treat classifi predict object closer shown linear combin embed base classifi natur correspond embed distanc measur base properti boostmap method reduc problem embed construct classic boost problem combin weak classifi optim strong classifi classif accuraci result strong classifi direct measur amount nearest neighbor structur preserv embed import properti boostmap embed optim criterion equal valid metric nonmetr space perform evalu databas hand imag handwritten digit time seri case boostmap significantli improv retriev effici small loss accuraci compar brute forc search boostmap significantli outperform exist nearest neighbor retriev method lipschitz embed fastmap vp tree index term index method embed method similar match multimedia databas nearest neighbor retriev nearest hierarch learn curv applic guidewir local fluoroscopi paper present method learn curv model detect segment close integr hierarch curv represent gener discrimin model hierarch infer algorithm appli method problem automat local guidewir fluoroscop sequenc fluoroscop sequenc guidewir appear visibl rigid dimension curv paper main contribut firstli present method learn complex shape appear free form curv hierarch model curv increas degre complex databas manual annot present comput paradigm context margin space learn algorithm close integr hierarch represent obtain fast paramet infer thirdli knowledg full system robustli local guidewir extens valid hundr frame present good quantit qualit result real fluoroscop video sequenc obtain frame queri dsensit embed common problem type databas retriev similar match queri object find match larg databas slow practic domain object compar computation expens similar distanc measur paper propos method approxim nearest neighbor retriev space method embed base mean construct function map object real vector space map preserv larg amount proxim structur origin space rapidli obtain short list match queri main novelti method construct embed queri sensit distanc measur measur distanc vector space term querysensit mean distanc measur depend current queri object report experi imag databas handwritten digit time seri databas case propos method outperform exist state art embed method mean significantli trade cienci retriev accuraci classmap effici multiclass recognit embed comput vision applic face recognit hand pose estim system recogn larg number class larg margin classif method adaboost svm provid competit accuraci rate cost evalu larg number binari classifi propos embed base method effici multiclass recognit method pattern class map vector pattern class tend map close test pattern small set candid class identifi effici simpl vector comparison experi hand pose recognit class face recognit class method time faster compar evalu binari classifi neglig loss classif accuraci queri dsensit embed common problem type databas retriev similar match queri object find match larg databas slow practic domain object compar computation expens similar distanc measur paper propos method approxim nearest neighbor retriev space method embed base mean construct function map object real vector space map preserv larg amount proxim structur origin space rapidli obtain short list match queri main novelti method construct embed queri sensit distanc measur measur distanc vector space term querysensit mean distanc measur depend current queri object report experi imag databas handwritten digit time seri databas case propos method outperform exist state art embed method mean significantli trade cienci retriev accuraci detect instanc shape class exhibit variabl structur paper propos method detect shape variabl structur imag clutter term variabl structur mean shape part repeat arbitrari number time part option part altern appear variat shape structur occur imag priori exist comput vision method includ deform model method design detect shape variabl structur detect shape decompos fix priori number part propos method handl variat shape structur variat appear individu shape part class shape model introduc call hidden state shape model natur repres shape variabl structur detect algorithm find instanc shape imag larg amount clutter find global optim correspond imag featur shape model experi real imag demonstr method local plant branch consist priori unknown number leav detect hand accur hand detector base chamfer distanc rotat invari index shape line draw effici nearest neighbor classif cascad approxim similar measur introduct nearest neighbor classifi appeal theirsimpl abil model wide rang complex parametr distribut find thenearest neighbor object larg databas long domain emploi comput alli expens distanc measur problem illustr comput alli expens imag similar measur call shape context match optic charact recognit athre nearest neighbor classifi shape context match train object yield error rate accur effici gestur spot prune subgestur reason abstract gestur spot challeng task locat start end frame video stream correspond gestur interest time reject gestur motion pattern paper propos gestur spot recognit algorithm base continu dynam program cdp algorithm run real time make gestur spot effici prune method propos system evalu small number hypothes compar cdp prune implement set model depend classifi learn train exampl make gestur spot accur subgestur reason process propos model fact gestur model fals match part longer gestur experi propos method prune subgestur model order magnitud faster accur compar origin cdp algorithm track analysi recognit human gestur video overview research autom gestur spot track recognit imag video comput group boston univers approach local track human hand video estim hand shape upper bodi pose track head facial motion effici spot recognit specif gestur video stream summar method effici dimension reduct gestur time seri boost classifi nearest neighbor search pose space model base prune gestur align hypothes algorithm demonstr domain american sign languag hand signal emploi flight director airport runwai gestur base interfac sever disabl user method gener appli domain requir effici detect analysi pattern time seri imag video onlin offlin charact recognit align prototyp nearest neighbor classifi simpl implement model complex parametr distribut provid state art recognit accuraci ocr databas time slow practic charact recognit reli similar measur requir computation expens pairwis align charact paper propos effici method comput approxim similar score charact base exact align small number prototyp propos method appli onlin offlin charact recognit similar base wide computation expens align method dynam time warp hungarian method case signific recognit speedup obtain expens minor increas recognit error queri dsensit embed common problem type databas retriev similar match queri object find match larg databas slow practic domain object compar computation expens similar distanc measur paper propos method approxim nearest neighbor retriev space method embed base mean construct function map object real vector space map preserv larg amount proxim structur origin space rapidli obtain short list match queri main novelti method construct embed queri sensit distanc measur measur distanc vector space term querysensit mean distanc measur depend current queri object report experi imag databas handwritten digit time seri databas case propos method outperform exist state art embed method mean significantli trade cienci retriev accuraci automat hand track video sequenc gestur sign languag video sequenc hand motion rapid hand frequent front front face hand locat ambigu naiv color base hand track insuffici improv track accuraci method emploi predict updat framework method requir care initi model paramet tend drift lose track extend sequenc paper tempor filter framework hand track propos initi reset human intervent frame simpl featur color motion residu exploit identifi multipl candid hand locat tempor filter viterbi algorithm select candid frame frame result track system automat identifi video trajectori unambigu hand motion detect frame track ambigu occlus overlap experi video sequenc hundr frame durat demonstr system abil track hand robustli detect handl track ambigu extract trajectori unambigu hand motion simultan local recognit dynam hand gestur method simultan local recognit dynam hand gestur propos core method dynam space time warp dstw algorithm align pair queri model gestur space time frame queri sequenc featur detector gener multipl hand region candid dynam program comput global match cost recogn queri gestur warp path align queri model sequenc time find hand candid region queri frame propos framework includ translat invari recognit gestur desir properti hci system perform approach evalu dataset hand sign digit gestur peopl wear short sleev shirt front background hand skincolor object algorithm simultan local gestur hand recogn hand sign digit dstw illustr gestur recognit set propos algorithm gener method match time seri multipl candid featur vector extract time step skin color dbase video segment time dvari illumin approach real time skin segment video sequenc approach enabl reliabl skin segment wide variat illumin track estim hand pose clutter imag method propos gener rank list plausibl dimension hand configur match input imag hand pose estim formul imag databas index problem closest match input hand imag retriev larg databas synthet hand imag contrast previou approach system function presenc clutter clutter toler index method computation effici approxim imag model chamfer distanc obtain embed binari edg imag high dimension euclidean space gener purpos probabilist line match method identifi line segment correspond model input imag occur chanc perform cluttertoler approach demonstr quantit experi hundr real hand imag databas index method hand pose estim estim hand pose gestur recognit applic rang human comput interact autom recognit sign languag paper hand pose estim treat databas index problem input imag hand similar imag larg databas hand imag retriev hand pose paramet retriev imag estim hand pose input imag lipschitz embed edg imag euclidean space improv effici databas retriev order achiev interact retriev time similar queri initi perform euclidean space paper describ ongo work focus choos refer imag order improv retriev accuraci hand pose reconstruct special map system recov hand pose monocular color sequenc propos system emploi linear supervis learn framework special map architectur sma map imag featur hand pose sma fundament compon set special forward map function singl feedback match function forward function estim directli train data case exampl hand joint configur visual featur joint angl data train set obtain cyberglov glove sensor monitor angular motion palm finger train visual featur gener comput graphic modul render hand arbitrari viewpoint joint angl viewpoint encod real valu real valu repres hand pose test system synthet sequenc sequenc color camera system automat detect track hand user calcul featur estim hand joint angl viewpoint featur result encourag complex task appear dbase framework hand shape classif camera viewpoint estim appear base framework hand shape classif simultan camera viewpoint estim present input imag segment hand similar match larg databas synthet hand imag retriev hand pose reconstruct special map system recov hand pose monocular color sequenc propos system emploi linear supervis learn framework special map architectur sma map imag featur hand pose sma fundament compon set special forward map function singl feedback match function forward function estim directli train data case exampl hand joint configur visual featur joint angl data train set obtain cyberglov glove sensor monitor angular motion palm finger train visual featur gener comput graphic modul render hand arbitrari viewpoint joint angl viewpoint encod real valu real valu repres hand pose test system synthet sequenc sequenc color camera system automat detect track hand user calcul featur estim hand joint angl viewpoint featur result encourag complex task estim predict evolv color distribut skin segment vari illumin approach real time skin segment video sequenc approach enabl reliabl skin segment wide variat illumin track explicit order markov model predict evolut skin color hsv histogram time histogram dynam updat base feedback current segment base predict markov model evolut skin color distribut frame parameter translat scale rotat color space consequ geometr parameter distribut propag warp sampl histogram paramet discret time dynam markov model estim maximum likelihood estim evolv time quantit evalu method conduct label ground truth video sequenc popular movi introduct locat track patch skin color fast 