optim kd dtree scalabl visual descriptor index learn combin multi dresolut spatial dweight doccurr matric imag represent dynam video collag imag search concept map imag search concept map interact imag search semant map tag refin regular lda summar tag imag collect cross dmedia repres vote linear neighborhood propag applic graph dbase semi dsupervis learn multipl label bayesian video search rerank find imag exemplar fast spars affin propag joint multi dlabel multi dinstanc learn imag classif normal tree partit imag segment maximum margin cluster pairwis constraint augment tree partit interact imag segment augment tree partit interact imag segment graph dbase semi dsupervis learn multi dlabel transduct video annot local learnabl kernel classifi optim video scene segment transduct multi dlabel learn video concept detect semi dsupervis classif universum universum data defin collect nonexampl belong class interest shown encod prior knowledg repres meaning concept domain problem hand paper address semi supervis classif problem call semi supervis universum simultan util label data unlabel data universum data improv classif perform propos graph base method make universum data depict prior inform classifi convent graph base semi supervis method graph regular util favor consist label propos method graph base easili extend multiclass case empir experi usp mnist dataset present show propos method obtain superior perform convent supervis semi supervis method semi dsupervis classif universum universum data defin collect nonexampl belong class interest shown encod prior knowledg repres meaning concept domain problem hand paper address semi supervis classif problem call semi supervis universum simultan util label data unlabel data universum data improv classif perform propos graph base method make universum data depict prior inform classifi convent graph base semi supervis method graph regular util favor consist label propos method graph base easili extend multiclass case empir experi usp mnist dataset present show propos method obtain superior perform convent supervis semi supervis method msra att trecvid high dlevel featur extract automat search joint affin propag multipl view segment joint segment simultan segment regist imag point reconstruct multipl view imag fundament structur data subsequ model applic paper treat joint segment weight graph label problem construct graph joint point joint similar measur propos hierarch spars affin propag algorithm automat jointli segment imag group point semi supervis affin propag algorithm propos refin automat result user assist final intens experi demonstr effect propos approach imag dbase tree model figur tree model set input photograph show exampl input result tree model imag inform foliag miss user sketch model approxim input imag form botan plausibl branch structur present method produc tree model input photograph limit user intervent approxim voxel base tree volum estim imag inform densiti valu voxel produc initi posit set particl perform flow simul particl trace tree basi combin form twig branch trunk order branch determin input photograph attractor particl simul geometri tree skeleton produc botan rule branch thick branch angl final leav ad initi seed particl simul lead varieti similar branch structur singl set photograph imag dbase model joint segment face recognit spectral featur face recognit spectral featur semi dsupervis classif linear neighborhood propag paper address gener problem learn label unlabel data base reason assumpt label data linearli reconstruct neighbor label develop approach call linear neighborhood propag lnp learn linear construct weight predict label unlabel data label data main procedur compos step comput local neighborhood geometri imag inform solv constrain squar problem obtain geometri predict label unlabel data solv linearli constrain quadrat optim problem lnp characterist firstli practic free paramet size neighborhood requir user tune easili extend sampl exampl final label achiev lnp suffici smooth respect intrins structur collect reveal label unlabel point promis experiment result object recognit data rank interact imag segment demonstr effect effici propos approach pictur collag paper address problem automat creat pictur collag group imag pictur collag kind visual imag summari arrang input imag canva allow overlai maxim visibl visual inform formul pictur collag creation problem bayesian framework salient region imag firstli extract repres set weight rectangl imag arrang formul maximum posterior map problem output pictur collag show visibl salient region overlaid imag effici markov chain mont carlo mcmc method design optim applic desktop imag brows imag search result summar demonstr effect approach imag dbase plant model creat model multipl view imag complex object challeng problem segment regist imag point reconstruct multipl view imag meaning object view fundament stage paper present joint segment framework imag base plant indoor model simultan segment point imag describ imag fragment method extract small leav imag base tree model experiment result demonstr segment critic imag base model visual object recognit probabilist kernel subspac similar multi dview em algorithm applic color imag segment paper propos algorithm multi view expect maxim algorithm multi view em deal real world learn problem natur split featur multi view em featur split train em success semi supervis learn algorithm text learn task consid multi view learn problem framework em algorithm multi view em algorithm impress advantag compar train em converg theoret guarante deal multipl view view util color imag segment discuss phenomenon weight color view coordin view lead segment result multi dview em algorithm applic color imag segment paper propos algorithm multi view expect maxim algorithm multi view em deal real world learn problem natur split featur multi view em featur split train em success semi supervis learn algorithm text learn task consid multi view learn problem framework em algorithm multi view em algorithm impress advantag compar train em converg theoret guarante deal multipl view view util color imag segment discuss phenomenon weight color view coordin view lead segment result probabilist tangent subspac unifi view tangent distanc td classic method invari pattern classif convent td pre obtain tangent vector difficult imag object paper extend td gener pattern classif task basic assumpt tangent vector approxim repres pattern variat propos probabilist subspac model encod variat linear subspac nonlinear subspac manifold subspac model model address unifi view probabilist tangent subspac pt experi show pt achiev promis classif perform imag data set kernel trick embed gaussian mixtur model abstract paper present kernel trick embed gaussian mixtur model gmm call kernel gmm basic idea emb kernel trick em algorithm deduc paramet estim algorithm gmm featur space kernel gmm view bayesian kernel method compar classic kernel method propos method solv problem probabilist framework tackl nonlinear problem tradit gmm avoid great comput cost problem exist kernel method larg scale data set emploi mont carlo sampl techniqu speed kernel gmm practic effici experiment result synthet real world data set demonstr propos approach satisf perform color imag segment kernel featur space 