Real-timeArticulatedHandPoseEstimationusingSemi-supervised
TransductiveRegressionForests
DanhangTang
ImperialCollegeLondon
London,UK
d.tang11@imperial.ac.uk
Tsz-HoYu
UniversityofCambridge
Cambridge,UK
thy23@cam.ac.uk
Tae-KyunKim
ImperialCollegeLondon
London,UK
tk.kim@imperial.ac.uk
Abstract
Thispaperpresentsthe rstsemi-supervisedtransduc-tivealgorithmforreal-timearticulatedhandposeestima-tion.Noisydataandocclusionsarethemajorchallengesofarticulatedhandposeestimation.Inaddition,thedis-crepanciesamongrealisticandsyntheticposedataunder-minetheperformancesofexistingapproachesthatusesyn-theticdataextensivelyintraining.WethereforeproposetheSemi-supervisedTransductiveRegression(STR)forestwhichlearnstherelationshipbetweenasmall,sparselyla-belledrealisticdatasetandalargesyntheticdataset.Wealsodesignanoveldata-driven,pseudo-kinematictech-niquetore nenoisyoroccludedjoints.Ourcontributionsinclude:(i)capturingthebene tsofbothrealisticandsyn-theticdataviatransductivelearning;(ii)showingaccura-ciescanbeimprovedbyconsideringunlabelleddata;and(iii)introducingapseudo-kinematictechniquetore near-ticulationsef ciently.Experimentalresultsshownotonlythepromisingperformanceofourmethodwithrespecttonoiseandocclusions,butalsoitssuperiorityoverstate-of-the-artsinaccuracy,robustnessandspeed.
(a)

RGB(b)

Labels(c)

Synthetic(d)Realistic
Figure1:Thering ngerismissingduetoocclusionsin(d),andthelittle ngeriswiderthanthesyntheticimagein(c).
1.Introduction
Articulatedhandposeestimationsharesalotofsimilar-itieswiththepopular3-Dbodyposeestimation.Bothtasksaimtorecognisethecon gurationofanarticulatedsubjectwithahighdegreeoffreedom.Whilelatestdepthsensortechnologyhasenabledbodyposeestimationinreal-time[2,24,12,26],handposeestimationstillrequiresimprove-ment.Despitetheirsimilarities,provenapproachesinbodyposeestimationcannotberepurposeddirectlytohandartic-ulations,duetotheuniquechallengesofthetask:
(1)Occlusionsandviewpointchanges.Selfocclusionsareprevalentinhandarticulations.Comparedwithlimbsinbodypose, ngersperformmoresophisticatedarticulations.Differentfrombodyposeswhichareusuallyupright

and
1
frontal[9],differentviewpointscanrenderdifferentdepthimagesdespitethesamehandarticulation.
(2)Noisyhandposedata.Bodyposesusuallyoccupylargerandrelativelystaticregionsinthedepthimages.Hands,however,areoftencapturedinalowerresolution.AsshowninFig.1,missingpartsandquantisationerroriscommoninhandposedata,especiallyatsmall,partiallyoccludedpartssuchas ngertips.Unlikesensornoiseanddeptherrorsin[12]and[2],theseartefactscannotbere-pairedorsmoothedeasily.Consequently,alargediscrep-ancyisobservedbetweensyntheticandrealisticdata.
Moreover,manuallylabelledrealisticdataareextremelycostlytoobtain.Existingstate-of-the-artsresorttosyntheticdata[16],ormodel-basedoptimisation[8,15].Nonethe-less,suchsolutionsdonotconsidertherealistic-syntheticdiscrepancies,theirperformancesarehenceaffected.Be-sides,thenoisyrealisticdatamakejointdetectiondif cult,whereasinsyntheticdatajointboundariesarealwayscleanandaccurate.
Addressingtheabovechallenges,wepresentanovelSemi-supervisedTransductiveRegression(STR)forest.Thisprocessisknownastransductivetransferlearning[21]:Atransductivemodellearnsfromasourcedomain,e.g.syntheticdata;ontheotherhand,itappliesknowledgetransformtoadifferentbutrelatedtargetdomain,e.g.real-isticdata,inthetestingstage.Asaresult,itbene tsfrom
Pictorial Structures Revisited People Detection and Articulated Pose Estimation(cvpr09)-slides_计算机软件及应用_IT/计算机_专业资料。广泛应用于行人检测和姿态...
ISR Real-time Articulated Hand Pose Estimation Danhang Tang*, Imperial 1685 using Semi-supervised College London Transductive Regression Forests Dahua Lin*, ...
Pictorial Structures Revisited People Detection and Articulated Pose Estimation(cvpr09)-slides_计算机软件及应用_IT/计算机_专业资料。广泛应用于行人检测和姿态...
热门文档