DynamicHairManipulationinImagesandVideos
MengleiChai
LvdiWang
YanlinWeng
XiaogangJin
KunZhou
StateKeyLabofCAD&CG,ZhejiangUniversityMicrosoftResearchAsia
(b)
(a)(c)(d)Figure1:Given(a)asingleportraitimageandafewuserstrokes,wegenerated(b)ahighquality3Dhairmodelwhosevisual delityandphysicalplausibilityenabledseveraldynamichairmanipulatingapplications,suchas(c)physically-basedsimulation,(d)combing,or(e,f)motion-preservinghairreplacementinvideo.OriginalimagescourtesyofAsianImpressionsPhotography.
Abstract
Thispaperpresentsasingle-viewhairmodelingtechniqueforgen-eratingvisuallyandphysicallyplausible3Dhairmodelswithmod-estuserinteraction.Bysolvinganunambiguous3Dvector eldex-plicitlyfromtheimageandadoptinganiterativehairgenerational-gorithm,wecancreatehairmodelsthatnotonlyvisuallymatchtheoriginalinputverywellbutalsopossessphysicalplausibility(e.g.,havingstrandroots xedonthescalpandpreservingthelengthandcontinuityofrealstrandsintheimageasmuchaspossible).Thelatterpropertyenablesustomanipulatehairinmanynewwaysthatwerepreviouslyverydif cultwithasingleimage,suchasdynamicsimulationorinteractivehairshapeediting.Wefurtherextendthemodelingapproachtohandlesimplevideoinput,andgeneratedy-namic3Dhairmodels.Thisallowsuserstomanipulatehairinavideoortransferstylesfromimagestovideos.
Keywords:hairmodeling,imagemanipulation,videoediting
1Introduction
Humanhairissodelicateand exiblethatpeoplealteriteasily,frequently,andinmanydifferentways,unlikeanyotherpartofourbody.Agentlebreeze,acasualbrush,oranelaboratehairdowillallreshapethehundredsofthousandsofstrandsintoadifferentarrangement.
Editinghairinaportraitimageis,however,farmoredif cult.With-outanyknowledgeoftheunderlying3Dstructure,traditionalim-ageeditingtoolscanhardlyperformmuchcomplexmanipulationonhair.Onereasonisthatintherealworldwhenhairisalteredevenslightly,somenewstrandsmaybeexposedwhilesomeoth-ersbecomeoccluded,resultinginanewimagewithnopixel-levelcorrespondencetotheoriginal.
Therecentprogresson3D-awareimagemanipulation[Hoiemetal.2005;Bitouketal.2008;Zhouetal.2010;Zhengetal.2012]hasdemonstratedthat,by ttingproper3Dproxiestotheobjectsof
interest,anumberofsemanticallymeaningfuloperationsthatarenearlyimpossibleinthe2Ddomainnowbecomeavailable.Asforhair,Chaietal.[2012]recentlyproposedamethodtogener-atea3Dhairmodelfromasingleportraitimage.Suchamodelvisuallymatchestheoriginalimagewhileretainingastrand-basedrepresentation,andthusenablessomeinterestingapplicationssuchas3D-awarehairreplacementorrenderingtheportraitfromanewviewpoint.However,themodelisbynomeansphysicallycorrect:thestrandsinitaremostlyshortsegments“ oating”inathinshelloftheentirehairvolume.Suchlimitationprecludesthepossibilityofmorecomplexanddynamichairmanipulatingoperationsbeingapplied.Forexample,tocorrectlysimulatetheeffectsofawindbloworahaircomb,thestrandrootsmustbe1) xedonthescalpandonlymovewiththeheadrigidly;Thestrandsthemselvesshouldalsobe2)smoothinaccordancewithhair’sphysicalpropertiesandnothavesharpbends;Moreover,theymust3)preservethelengthandcontinuityoftherealstrandsintheimageasmuchaspossible,sothatwhenlocallydeformed,theyarestillcontinuousstrandsin-steadofdisconnectedfragments.Wecallamodelsatisfyingthesethreeconditionsaphysicallyplausiblehairmodel.
Astraightforwardwaytoachievephysicalplausibilityisto“grow”eachstranddirectlyfromthescalpsurfaceandfollowasmoothlychangingdirectionestimatedfromtheimage.Unfortunately,strandsgeneratedlikethismaynotcorrespondwelltotherealonesintheimageduetothelimitedaccuracyofdirectionestimationandotherimageartifacts.Morespeci cally,localimageorienta-tionestimatorsalonecannotdiscriminatehair’sdirectionofgrowth.Withoutresolvingsuchdirectionalambiguity,strandswithoppositedirections(e.g.,thosenearahairparting)couldbemistakenlycon-nected.Tomakethingsworse,suchana¨ vehairgrowingprocessdoesnotevenguaranteefullcoverageofthehairregion,i.e.,someregioncontaininghairintheimagewillneverbereachedbyastrandgrownfromthescalp.
Inordertofacilitatedynamichairmanipulationinimages,wemustgenerateamodelthatisbothvisuallyandphysicallyplausible.
1.1Contributions
Thisistheauthorsversionofthework.ItispostedherebypermissionofACMforyourpersonaluse.Notforredistribution.Thede nitiveversionwaspublishedinACMTransactionsonGraphics,32,4,July2013.
Inthispaperwehavedevelopedanovelsingle-viewhairmodelingalgorithm(§2).Givenaportraitimageandafewuserannotationsasinput,itgeneratesahigh-qualitystrand-based3Dhairmodelthat
Dynamic manipulation w... 暂无评价 8页 免费 Dynamic manipulation w... ...? Hair roots are fixed on the scalp Preserve length and continuity ? ...
(Guest Editors) Reanimating Faces in Images and ...from snapshots of dynamic sequences 27 18 25 . ...Hair was covered by a bathing cap. 17 scans ...
Pose primitive based human action recognition in videos or still images ...or dynamic features and thus allows for action recognition in still images....
Salt and pepper noise is a typical type of noise in images and videos. ...and minimum values in the dynamic range (0, 255) (Ratnababu et al., ...
important topics related to the creation and manipulation of digital images ?... and light field rendering –––––– achieve high dynamic range extended...
DYNAMIC VISION From Imag... 46页 免费 Facial Action Recognitio... 13页... Face Recognition from Still Images and Videos In most situations, ...
For modeling di?erent textures in videos ...of dynamic textures by manipulation of the model.... be a sequence of τ images and let y(t) =...
High dynamic range video using split aperture camera_专业资料。We present ...[3] for corresponding images in the videos (We will not discuss the ...
We present and formulate various policies used for the effective mapping of videos in the runtime of the TV-Anytime system. A predictive隐藏>> Dynamic ...
and dark surfaces such as a black hair region, because reflected lights S...images captured by a TOF depth sensor to generate a dynamic 3D scene with...
热门文档