Temlae-Based Reconsrucon of Surface Mesh Anmaon from Pon Cloud Anmaon Sang Il Park and Seong-Jae Lm In hs aer, we resen a mehod for reconsrucng a surface mesh anmaon sequence from on cloud anmaon daa. We manly focus on he arculaed body of a subec he moon of whch can be roughly descrbed by s nernal skeleal srucure. The on cloud daa s assumed o be caured ndeendenly whou any ner-frame corresondence nformaon. Usng a emlae model ha resembles he gven subec, our basc dea for reconsrucng he mesh anmaon s o deform he emlae model o f o he on cloud (on a frame-by-frame bass) whle mananng ner-frame coherence. We frs esmae he skeleal moon from he on cloud daa. Afer alyng he skeleal moon o he emlae surface, we refne o f o he on cloud daa. We demonsrae he vably of he mehod by alyng o reconsruc a fas dancng moon. Keywords: Pon cloud anmaon, mesh anmaon, characer anmaon. Manuscr receved Nov. 25, 2013; revsed June 2, 2014; acceed June 24, 2014. The research was suored by Basc Scence Research Program hrough he Naonal Research Foundaon of Korea (NRF) funded by he Mnsry of Educaon, Scence and Technology (2012 R1A1A1014702), and hs work was suored by he IT R&D Program of MSIP/IITP (10047093, 3D Conen Creaon and Edng Technology Based on Real Obecs for 3D Prnng). Sang Il Park (sark@seong.ac.kr) s wh he Dearmen of Dgal Conens, Seong Unversy, Seoul, Re. of Korea. Seong-Jae Lm (corresondng auhor, slm@er.re.kr) s wh he SW Conen Research Laboraory, ETRI, Daeeon, Re. of Korea. I. Inroducon The acquson of dynamcally varyng geomery daa from a deformng obec has been an acve research area n comuer grahcs. Convenonal mehods o caure such a me-varyng shae are mosly based on vson echnques. Usng synchronzed mulle cameras n a conrolled sudo, he 3D surface daa a each nsance are caured by exlong varous sereo machng mehods. The deals of he rocess can dffer accordng o he choce of hardware seus and mlemenaons. However, he reresenaon of her resulng daa usually shares he same form, whch s he hreedmensonal on cloud. The on cloud has rmarly been used for caurng a sac obec. However, also faclaes he reresenaon of anmaon daa by concaenang he resulng on clouds, caured ndvdually a each frame, n emoral order. Alhough he on cloud reresenaon has roven useful for sorng and renderng wh remendous geomerc deals, s orgnally nended o be used for a sac obec. When comes o a dynamc obec, several ssues may arse for examle, he comresson of large-sze daa for effcen memory use and he consrucon of ner-frame emoral coherence for effecve os-rocessng (ose-edng or reexurng). On he oher hand, olygonal surface reresenaons, such as rangular meshes, have been more oular for anmaon uroses, n whch an anmaon s reresened by he raecores of verces whle keeng he connecvy beween such verces unchanged. Moreover, here are already many research resuls avalable dealng how o effecvely comress and manulae such mesh anmaon daa [1] [2]. Thus, he oon o use on cloud daa for anmaon uroses s 1008 Sang Il Park and Seong-Jae Lm 2014 ETRI Journal, Volume 36, Number 6, December 2014 h://dx.do.org/10.4218/er.14.0113.1181
rovng ncreasngly more aracve. In hs aer, we resen a mehod for converng on cloud anmaon daa no mesh anmaon daa wh consan connecvy. We ake he on cloud anmaon sequences whou he ner-frame coherency as an nu. Our roblem s dfferen from he revous work [3] [4] for generang mesh anmaon from mages caured from synchronzed mulle vdeo cameras, n whch s assumed ha he slhouee nformaon can be exraced easly whou much nose. In our case, however, he on cloud daa are already rocessed and can conan severe nose esecally around he regon of he occluson, whch s no easly searable. Thus, nose-robus reconsrucon s he man ssue o be addressed n our mehod. In addon, he full-body on cloud anmaon daa ofen lacks deals due o he lmed number of full-body-sho cameras. I s, herefore, desrable o augmen he deals of he caured daa. Our basc dea for a robus and conssen ransformaon s o use a emlae model. The emlae model s no necessarly o be he same subec o be caured. However, he nernal skeleal srucure s assumed o be he same. The use of a emlae model resuls n several benefs: frs, he emlae model enables us o easly buld a emoral coherence by regserng s geomerc elemens o he unsrucured on clouds a each frame. Second, we can archve more deals o he orgnal caured daa. Fnally, he emlae model hels o remove he noses exsng n he orgnal nu daa. We address hree man ssues: frs, we resen a mehod o adus he gven general emlae model o beer f he geomery of a arcular subec. Second, we recover he moon of he subec and ransform no a mesh anmaon. Fnally, we reduce he vsual arfacs of he obaned mesh anmaon by alyng a local deformaon fler. The remander of he aer s organzed as follows: we frs dscuss relaed work n Secon II. We hen brefly gve an overvew of our mehod n Secon III. The mehod consss of hree man ars, each of whch s descrbed n deal n Secon IV. In he las wo secons, we rovde our exermenal resuls and dscuss he advanages and lmaons of he mehod. II. Relaed Work Caurng anmaon daa from a deformng obec has been an moran roblem n comuer anmaon for many years. Here, we brefly summarze he work mos relaed o our own by manly focusng on he dynamc full-body deformaon acquson. The echnques can be roughly caegorzed no wo classes: emlae-based mehods and non-emlae-based mehods. Our mehod falls no he former caegory. Temlae-based mehods exlo a emlae model as a ror of he subec o be caured. Temlae models have been wdely used mosly for esmang he corresondences and for augmenng mssng nformaon n he caured daa. The frs full-body dynamc caure of human subecs was done by Sand and ohers, n 2003 [5]. Alhough hey dd no use a geomery model as a emlae, a redefned skeleal srucure ogeher wh a rmve deformaon model was gven o asss n he esmaon of he confguraon of he subecs. Allen and ohers bul a general emlae model o esmae he deformaon sace wh resec o dfferen body shaes [6]. In her work, hey defned a small se of redefned landmarks on he emlae model for easly esmang corresondences beween caured daa. For he same urose, he redefned landmarks have been commonly used n oher mehods, such as hose n [7] [9]. However, hey were mosly for acqurng sac obecs or a few sequences of dynamc obecs, due o he comlexy of her auomac mehod of buldng corresondences. In 2008, wo neresng works were resened smulaneously by wo dfferen grous; her mehods havng smlar caure seus. One was by Vlasc and ohers [4], whle he oher was by de Aguar and ohers [3]. Boh use emlae models exensvely for corresondences and fllng mssng nformaon. Esecally, [3] exloed a lower resoluon verson of he volumerc emlae model for effecvely rackng fas and comlex nonrgd moon. Mos recenly, L and ohers resened a mehod for caurng a comlex dynamc moon only n a snglevew seu [10]. They also used her emlae a a lower resoluon so as o acheve robusness and effcency of caures. Non-emlae-based mehods do no use models or redefned corresondences. Mra and ohers suggesed o use a se of frames wh dense saal and emoral daa drecly o comue he moon of a scanned obec [11]. In [12], Sussmuh and ohers comued a four-dmensonal mlc surface aroxmang he nu on cloud anmaon and reconsruced a olygonal mesh anmaon n an as-rgd-as-ossble manner. Wand and ohers used a deformable machng based on a sascal omzaon for he smulaneous esmaon of he shae and he moon [13]. In her followng work [14], hey mroved s effcency by searaely handlng he shae and he moon n he omzaon. Mos recenly, Tevs and ohers resened a mehod ha frs deeced a small se of landmarks commonly shown n he gven on cloud sequences and hen exended hem o fnd a dense se of corresondences [15]. Those mehods usually assume ha he on cloud daa have enough deals wh less nose. However, our daa lack deals and conan many noses. ETRI Journal, Volume 36, Number 6, December 2014 Sang Il Park and Seong-Jae Lm 1009 h://dx.do.org/10.4218/er.14.0113.1181
III. Overvew We ake he sequence of he on cloud from a non-rgdly deformng subec as nu daa, as shown n Fg. 1, ogeher wh a olygonal surface model as a emlae. Because a human subec s our man neres, we use a general humansurface model wh 4,000 verces, commercally avalable n [16], for he emlae. We mgh exlo a hree-dmensonal scanner o ge he exac, dealed geomery from he subec. However, here, we consder a general scenaro n whch only he dynamcally caured daa are avalable. The emlae surface model s re-rocessed so as o have lnear blend sknnng (LBS) weghs for each verex; hence, an arculaed deformaon can be aled. Fgure 2(c) shows our nal surface model and he assgned LBS weghs, where we segmen he whole body no 17 body ars. The on cloud daa conss of a large number of ons. For examle, he average number of ons er frame s abou 32,000 n our daa se. Dealng wh such a large number of ons s no effcen when aemng o erform manulaons on hem. Thus, we rocess he on cloud daa ndvdually a each frame so ha can be reresened n a mul-resoluon manner. Ths rocess s moran, oherwse he followng rocesses would no be feasble because of he requred comuaon me. To generae surface anmaon wh he gven daa, our mehod consss of wo searae rocesses: one for he frs frame, and he oher for he remanng frames, as llusraed n Fg. 3. We frs regser he emlae surface o he on cloud daa a he frs frame. For an effecve regsraon, he subecs were asked o sar each caure sesson wh a ose smlar o ha of he emlae model. Addonally, we rovde ars of corresondng feaures beween he emlae and he on cloud. In our exermens, we defne 44 corresondences on he facal feaures, such as nose, eyes, and mouh, and boney landmarks of he body, such as knees, elbows, shoulders, and so on. Afer he frs-frame regsraon, we sequenally generae a machng surface model o he on cloud a a gven frame by deformng he surface model obaned a he revous frame. Fg. 1. Excers from he nu on cloud anmaon daa of he shuffle dance. Each frame consss of abou 32,000 ons. (a) (b) (c) (d) Fg. 2. Frs-frame regsraon: (a) on cloud daa a frs frame wh 34,720 ons, (b) reduced on cloud daa wh 2,000 ons, (c) gven emlae surface model, and (d) regsered surface model of frs frame. Pon cloud anmaon Frs-frame regsraon Pon cloud a he 1s frame Regsraon Temlae surface model wh LBS weghs Consecuve-frame regsraon Pon cloud daa a each frame Skeleal moon esmaon Surface refnemen Fg. 3. Overvew of sysem. Surface mesh anmaon Ths er-frame regsraon consss of wo sages: we frs esmae he skeleal srucure by comung he on angles resemblng he ose of he gven daa. The skeleal nformaon s used n he alcaon of he LBS weghs o he emlae model for he generaon of an nal esmae o he gven frame. Then, we fnd he bes er-verex deformaon of he surface n whch he dfference beween he on cloud and he surface model s mnmzed whle keeng he smoohness of he surface. Alhough he emoral coherence beween consecuve frames s loosely consdered n he skeleon esmaon sage, here sll exs arfacs of he emoral nconssency n he resulng mesh anmaon. By exlong he esmaed skeleal moon, we reduce he arfacs n a os rocess whle reflecng he non-rgd deformaon. IV. Mehod 1. Daa Reresenaon The on cloud anmaon daa conss of a sequence of he on cloud. Each on cloud s caured ndvdually a each frame and ncludes a large number of ons. Among hese ons, here exss neher me coherence beween frames nor saal relaonsh, such as neghborng nformaon a a frame. We denoe he on cloud a a gven frame as a se,,..., P 1 2, where N N s he number of ons and s he oson of he h on. The average value of N n our exermens s abou 32,000. For esmang basc 1010 Sang Il Park and Seong-Jae Lm ETRI Journal, Volume 36, Number 6, December 2014 h://dx.do.org/10.4218/er.14.0113.1181
surface roeres, such as a normal and curvaure, we sore he ndces of he closes K neghbors for each on based on he Eucldan dsance. Such a large number of ons slow down basc oeraons on he ons, such as searchng. I also makes hard o reflec global feaures. Thus, for beer effcency and effecveness, we consruc a searae lower-resoluon verson of he daa n a mul-resoluon manner based on he mehod of [17] as follows: we frs cluser a se of ons such ha hey are close o each oher as well as close o he 3D lane aroxmang he on dsrbuon n he cluser. In hs way, we can ge a bgger cluser on he lanar regon and a smaller cluser around a hgher curvaure regon. Then, by ckng he on, for each cluser, closes o s cener as a reresenave, we buld a lower-resoluon verson of he on cloud. We denoe hs ˆ ˆ, ˆ,..., ˆ P 1 2 Nˆ wh smaller verson a frame as ˆ N ons. Our average value of N ˆ s abou 2,000. Fgure 2(b) shows he reduced on cloud a he frs frame. We use he lower-resoluon verson when a rough esmae s adequae, such as n he early sages of he omzaon, and use he full verson when greaer accuracy s requred. The emlae surface model M s comosed of he verces and her connecvy nformaon. For smlcy of exlanaon, he emlae model s assumed o be a rangular mesh. Then, can be reresened as a se of verex osons and a se of verex ndex rles reresenng a rangle. We denoe he oson of he h verex as v, 1 N v, where N v s he number of verces n he mesh, and a rle for 1 2 3 rangle as,,. Each verex s assgned wh a se of LBS wegh values, s,, 1 s Ns, where s, s he wegh value of verex for body ar s, N s s he oal N number of body ars, and s 1, 1 s s. In our exermens, N s s 17, as shown n Fg. 2(c), where dfferen colors are used o ndcae dfferen ars. 2. Frs-Frame Regsraon We regser he emlae surface model M o he on cloud P 1 a he frs frame by deformng he surface. We ado he omzaon framework of Allen and ohers [2]: we frs fnd he global ransformaon of M o have he bes f wh P 1. The global ransformaon s defned as a combnaon of ranslaon and roaon followed by scalng. Afer he global machng, we omze he local deformaon a each verex. We reresen he local deformaon a each verex as a 3 4 affne ransformaon marx, whch we denoe as A for verex. For an effecve machng, we assgn ar-wse feaure corresondences beween seleced verces n M and ons n P 1. We denoe he se of he feaure verex ndces of he surface M P and s corresondng se for he on cloud as K and K, resecvely, where 1 N c, and N c s he oal number of corresondences. We selec 44 feaure ars for he corresondence n our exermens. Noe ha hs assgnmen s done us once for he frs frame. The global machng s erformed by consderng only hese corresondng ars. The global ranslaon s easly found by comung he dslacemen beween he cener of he chosen verces and ha of he corresondng ons. The global roaon s obaned by usng he mehod of Horn [18] for he absolue orenaon roblem. Fnally, we aly he leas squares fng o fnd he global scalng value. We denoe he oson of verex afer he global ransform as v~. The omzaon for he local deformaon s o fnd he bes affne ransform of every verex ha mnmzes he combnaon of he followng hree error erms: he frs s he daa error E d for measurng gas beween all he ons and he surface, he second s he smoohness error E s for keeng he orgnal deals of he surface, and las s E f, whch s for reflecng he user-gven corresondng feaures durng omzaon. Please noe ha we mosly follow he conce of he hree error erms from [6]. However, we cusomze he daa error erm o be alcable o our on cloud daa. We gve a dealed descron of he aforemenoned error erms n he followng subsecons. A. Daa Error The man urose of he regsraon s o make he shae of he emlae model resemble he gven on cloud. We can measure he closeness of hs machng by comung he dsances from each on o he closes surface. Because our emlae model s a rangular mesh, we frs fnd he closes rangle o each on from he on cloud. For he closes rangle c o he on 1, we can also comue he closes oson on he rangle and reresen n s barycenrc 1 2 3 coordnae wc, w, c w c wh he hree verces of he rangle. Then, he daa error becomes E 1 N 3 2 1 k d wc k k c c 1 k1 A v, 1 where N s he number of he ons n he on cloud a he frs frame. B. Feaure Error Ths error s desgned o reduce he dsance beween he corresondng ars, whch gudes he resulng surface o have a conexual machng o he on cloud. Ths erm s ETRI Journal, Volume 36, Number 6, December 2014 Sang Il Park and Seong-Jae Lm 1011 h://dx.do.org/10.4218/er.14.0113.1181
esecally necessary when he shae of boh he emlae and he on cloud are no nally close enough o each oher. By denong he verex ndex of he h ar of he M P corresondences as K and he ndex of he on as K, he error s defned as follows: E C. Smoohness Error Nc 1 f A M v M P K K K 1 2. The local deals can be reserved f a verex and s conneced neghborng verces undergo a smlar local deformaon o each oher. Thus, we measure he local change by comarng he dfference n he affne ransform marces of he neghborng verces as follows: E Nv d 2 s A A, k F, 1 k1 where s he Frobenus norm, A F k, s he affne marx of he kh 1-rng neghbor of verex, and d s he number of he 1-rng neghbors. D. Omzaon Consderng he above hree error erms, he oal error E s he weghed combnaon of hem. Tha s, E E E E, d d f f s s where d, f, and s are he resecve wegh values. We run he omzaon based on he conugae graden mehod. Inally, we assgn weghs only o f and s so ha he model s regsered globally frs. Afer he convergence, we erform he omzaon agan wh equal weghs for all hree erms. 3. Consecuve-Frame Regsraon We use he regsered model for he frs frame as he new emlae model for he res of he frames because he model s already secfc o he subec. Our regserng rocess consss of hree subrocesses. We frs esmae he skeleal moon for all he frames, whch rovdes a good nal esmae for he surface. Then, we omze he local deformaon o f o he on cloud daa. Fnally, we reduce he nose from he resulng surface anmaon by usng ner-frame coherency. A. Skeleal Moon Esmaon From he frs frame regsraon, we esmae he skeleal srucure of he subec, ncludng he on osons and lengh of bones. Because he skeleal moon s used for he rough nal esmae and for he laer local regsraon, s no requred o be very accurae. For hs reason, we assume ha he skeleal srucure s aroxmaed from he gven LBS wegh values, by whch he number of bones s he same as he number of body ars, and whereby he on osons are locaed a he cener of he boundary beween he neghborng body ars. Whle keeng he lengh of he bone consan, we esmae he on confguraon for each frame, whch s arameerzed wh he oson of he roo on and roaons of all he ons. Sarng from he second frame, we sequenally fnd he bes on confguraon by alerng one on from he revous frame such ha afer alyng LBS o he emlae surface, he dsance beween he on cloud and he surface should be mnmzed whle keeng a mnmal change o he revous frame. We exlo a graden-based mehod for he omzaon. B. Surface Refnemen Wh he skeleon ose a a gven frame, we frs aly LBS o he emlae surface model o ge he nal surface for he omzaon. Then, we fnd he bes local deformaon by mnmzng he error erm, n a manner smlar o ha gven n Secon IV-2, wh he exceon of he corresondence erm, whch s only avalable a he frs frame. C. Pos Nose Reducon E E E d d. s s Our rocess of regsraon s bascally a er-frame rocess, n whch each frame s consdered ndeendenly exce for he esmaon of he skeleal moon. Thus, nevably can lose ner-frame conssency, whch resuls n a erkness n he moon of each verex. Vsually, hs severely degrades he qualy of he resulng anmaon. To remedy he arfacs, we resen a nose reducon rocess. From our observaons, he arfacs are more obvous when he moon of a verex s dfferen from ha of s neghborng verces. Borrowng an dea from mul-resoluon sgnal rocessng, we dvde he moon of a verex no wo levels: one s he global moon, and he oher s he local moon. Because of he assumon of he arculaed body for he subec, we defne he global moon for each verex as he moon resulng solely from he LBS-weghed skeleal moon. Then, he local moon can be defned as he dfference beween he comued moon and he global moon. Gven he oson v of he h verex of he resulng surface a frame, we denoe he oson of he verex obaned by only alyng he skeleal moon o he emlae model as v. global 1012 Sang Il Park and Seong-Jae Lm ETRI Journal, Volume 36, Number 6, December 2014 h://dx.do.org/10.4218/er.14.0113.1181
Then, he local oson local v s comued as follows: local global v v v. Because he global moon s smooh due o he smooh skeleal moon, we erform he emoral smoohng oeraon on he local oson. We use a Gaussan fler wh a kernel sze of fve frames. By dong hs, we kee he global moon unchanged and reduce he locally abru movemen of he verces. V. Exermenal Resuls We mlemened he algorhm usng C++ wh Wndows 7. The exermens were erformed on a PC wh an Inel 7 CPU a 3.2 GHz wh 8 GB of memory. We aled our mehod o reconsruc he fas dancng moon of he acor shown n Fg. 1. The oal number of frames s 300, and each frame consss of abou 35,000 ons. Our emlae surface model, as shown n Fg. 2, s made u of 4,000 verces. The resulng mesh anmaon s gven n Fg. 4. As exlaned n Secon IV, we frs esmae he skeleal moon, as shown n he mddle row of Fg. 4. Afer alyng he skeleal moon o he emlae surface, we refne o f o he on cloud, as shown n he Fg. 5. Close-u vew. Inu on cloud daa conssng of nose due o occlusons durng he caure. However, our reconsrucon reduces he nose hrough he use of a emlae model. (mm) 7 6 5 4 Average dsance beween on cloud and regsered surface 3 0 50 100 150 200 250 Frames Fg. 6. Reconsrucon error. boom row of Fg. 4. The average rocessng me aken for each frame s abou 20 seconds. The resulng move fles can be found a h://dasan.seong.ac.kr/~sark/oncloud/. Because of he use of a emlae model, our mehod can reduce he noses exsng n he orgnal daa, orgnang from he vsual occlusons, durng he caure sesson. As shown n Fg. 5, here exs arfacs near he lef arm, whch are removed n our reconsruced mesh anmaon. Addonally, he emlae surface model augmens he deals o he reconsruced anmaon. For examle, our resulng mesh anmaon has dealed geomeres around he fngers and he ears. Fgure 6 shows he reconsrucon error for all frames. The error s measured as he average dsance beween he on cloud and he reconsruced surface. The errors range from 3.7 mm o 6.2 mm. Fg. 4. Our exermenal resuls. To row: nu on cloud daa; mddle row: esmaed skeleal oses; boom row: reconsruced mesh anmaon. VI. Dscusson In hs aer, we resen a mehod for converng on cloud ETRI Journal, Volume 36, Number 6, December 2014 Sang Il Park and Seong-Jae Lm 1013 h://dx.do.org/10.4218/er.14.0113.1181
anmaon daa no mesh anmaon daa. Our man dea s o use a emlae surface model wh an nernal skeleal srucure for rackng and esmang he me-varyng geomery of he arculaed subec. We demonsrae he vably of he mehod by alyng o a fas dancng moon. Our mehod has several lmaons. Frs, our mehod sequenally fnds he skeleal moon by advancng he frames one a a me. Thus, f any falures haen n a frame, hen he error can be accumulaed n he followng frames. To fx hs roblem, one ossble soluon s o allow he user o adus he on confguraon neracvely durng he rocess, as Vlasc and ohers dd n [4]. Second, he smoohness error erm n he omzaon may reven local deformaons of he surface. Ths s fne when nose exss n he nu daa. However, can also remove desred deformaons, such as wrnkles on clohes. Fnally, we assume ha he emlae mesh and s skeleal srucure are gven by he user. However, for more usably, would be desrable f he sysem could esmae hose daa from he on clouds so ha can be aled o an arbrary subec; for examle, anmals wh dfferen skeleal srucures [19]. References [1] D.L. James and C.D. Twgg, Sknnng Mesh Anmaons, ACM Trans. Grah., vol. 24, no. 3, July 2005,. 399 407. [2] S. Krcher and M. Garland, Edng Arbrarly Deformng Surface Anmaons, ACM Trans. Grah., vol. 25, no. 3, July 2006,. 1098 1107. [3] E. de Aguar e al., Performance Caure from Sarse Mulvew Vdeo, ACM Trans. Grah., vol. 27, no. 3, 2008,. 98:1 98:10. [4] D. Vlasc e al., Arculaed Mesh Anmaon from Mul-vew Slhouees, ACM Trans. Grah., vol. 27, no. 3, 2008,. 97:1 97:9. [5] P. Sand, L. McMllan, and J. Poovc, Connuous Caure of Skn Deformaon, ACM Trans. Grah., vol. 22, no. 3, July 2003,. 578 586. [6] B. Allen, B. Curless, and Z. Poovc, The Sace of Human Body Shaes, ACM Trans. Grah., vol. 22, no. 3, July 2003,. 587 594. [7] D. Anguelov e al., Recoverng Arculaed Obec Models from 3D Range Daa, Proc. Conf. Uncerany Arf. Inell., Banff, Canada, 2004,. 18 26. [8] D. Anguelov e al., Scae: Shae Comleon and Anmaon of Peole, ACM Trans. Grah., vol. 24, no. 3, July 2005,. 408 416. [9] A.M. Bronsen, M.M. Bronsen, and R. Kmmel, Generalzed Muldmensonal Scalng: A Framework for Insomery-Invaran Paral Surface Machng, Proc. Naonal Academy Sc., vol. 103, no. 5, 2006,. 1168 1172. [10] H. L e al., Robus Snglevew Geomery and Moon Reconsrucon, ACM Trans. Grah., vol. 28, no. 5, Dec. 2009,. 174:1 175:10. [11] N.J. Mra e al., Dynamc Geomery Regsraon, Proc. Eurograhcs Sym. Geomery Process., Barcelona, San, 2007,. 173 182. [12] J. Sussmuh, M. Wner, and G. Grener, Reconsrucng Anmaed Meshes from Tme-Varyng Pon Clouds, Proc. Sym. Geomery Process., Coenhagen, Denmark, 2008,. 1469 1476. [13] M. Wand e al., Effcen Reconsrucon of Non-rgd Shae and Moon from Real-Tme 3D Scanner Daa, ACM Trans. Grah., vol. 28, no. 2, Ar. 2009,. 15:1 15:15. [14] M. Wand e al., Reconsrucon of Deformng Geomery from Tme-Varyng Pon Clouds, Proc. Sym. Geomery Process., Barcelona, San, 2007,. 49 58. [15] A. Tevs e al., Anmaon Carograhy Inrnsc Reconsrucon of Shae and Moon, ACM Trans. Grah., vol. 31, no. 2, Ar. 2012,. 12:1 12:15. [16] CGHuman, CGCharacer, 2013. Accessed Aug. 26, 2013. h://www.cgcharacer.com/cghuman.hml [17] M. Pauly, L. Kobbel, and M.H. Gross, Mul-resoluon Modelng of Pon-Samled Geomery, ETH CS Techncal Reor, no. 373, 2002. [18] B. Horn, Closed-Form Soluon of Absolue Orenaon Usng Un Quaernons, J. O. Soc. Amerca, vol. 4, no. 4, Ar. 1, 1987,. 629 642. [19] M. Sung, Fas Moon Synhess of Quadruedal Anmals Usng a Mnmum Amoun of Moon Caure Daa, ETRI J., vol. 35, no. 6, Dec. 2013,. 1029 1037. 1014 Sang Il Park and Seong-Jae Lm ETRI Journal, Volume 36, Number 6, December 2014 h://dx.do.org/10.4218/er.14.0113.1181
Sang Il Park receved hs PhD degree n comuer scence from he Korea Advanced Insue of Scence and Technology, Daeeon, Re. of Korea, n 2004. Afer workng for abou wo years as a osdocoral fellow a Carnege Mellon Unversy, Psburgh, PA, USA, and hen a he Naonal Insue of Advanced Indusral Scence and Technology, Tokyo, Jaan, he s currenly afflaed wh he Dearmen of Dgal Conens, Seong Unversy, Seoul, Re. of Korea, as an asssan rofessor. Hs rmary research neress are synheszng characer anmaon and vsually smulang naural henomena. Seong-Jae Lm receved hs PhD degree n nformaon and communcaon engneerng from he Gwangu Insue of Scence and Technology, Gwangu, Re. of Korea, n 2006. From 2004 o 2005, he was wh he Unversy of Pennsylvana, Phladelha, USA, as a vsng scholar a he Medcal Image Processng Laboraory. He s currenly a senor researcher of he Creave Conen Research Laboraory, Elecroncs and Telecommuncaons Research Insue, Daeeon, Re. of Korea. Hs rmary research neress are sensor-based human modelng, deformaon, and anmaon. ETRI Journal, Volume 36, Number 6, December 2014 Sang Il Park and Seong-Jae Lm 1015 h://dx.do.org/10.4218/er.14.0113.1181