06.08.2013 Views

Paper - Geometric Algorithms for Modeling, Motion, and Animation ...

Paper - Geometric Algorithms for Modeling, Motion, and Animation ...

Paper - Geometric Algorithms for Modeling, Motion, and Animation ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

4.1 Implementation<br />

Wehaveimplementedouralgorithmonast<strong>and</strong>ard2.4GHzIntel<br />

Pentiummachinewith4GBRAMon32-bitWindows/XPplat<strong>for</strong>m.<br />

Theper<strong>for</strong>manceismeasuredusingasinglethread.Weusek-DOPs<br />

(specifically16-DOPs)asboundingvolumesbecausetheyprovide<br />

agoodbalancebetweentightfitting<strong>and</strong>rapidupdating.<br />

Weuserestructuring<strong>and</strong>refittingtoupdatethehierarchy<strong>for</strong>de<strong>for</strong>mablemodels.IntelSSE/SSE2instructionsareusedtoacceleratetheupdatingof<strong>and</strong>overlaptestingbetweenboundingvolumes.<br />

WeuseanimplementationoftheICCDalgorithmbasedonnormal<br />

cones<strong>for</strong>high-levelculling[Tangetal.2009a]asabaseline<strong>for</strong><br />

comparison.WealsointegrateourfilterswiththeR-Trianglealgorithm[Curtisetal.2008],asitspendsalargefractionofthequery<br />

timeontheelementarytests.<br />

4.2 Benchmarks&Per<strong>for</strong>mance<br />

Inordertotesttheper<strong>for</strong>manceofouralgorithm,weusedsixdifferentbenchmarks,arisingfromdifferentsimulationswithdifferent<br />

characteristics.<br />

• Lion:Forthisbreakingbenchmarkwith1.6Mtriangles(Figure1),ouralgorithmreducesthenumberofelementarytest<br />

by 10xascomparedtopriorCCDalgorithms.<br />

• Balls:Ascenewithhundredsofballs(34Ktriangles)thatare<br />

collidingwitheachother(Figure7(a)).Ourde<strong>for</strong>mingnonpenetrationfilterreducesthenumberofelementarytestsby<br />

17x.<br />

• FallingAlphabets: Multiplede<strong>for</strong>mingcharacters(5Ktriangles)fallintoabowl<strong>and</strong>breakintopieces(Figure7(b)).<br />

OurCCDalgorithmreducesthenumberofelementarytests<br />

by 58x.<br />

• Princess: Thismodel(40Ktriangles)hasmanyinter-<strong>and</strong><br />

intra-objectcollisions(Figure7(c)). Ouralgorithmreduces<br />

thenumberofelementarytestsby 20.5x.<br />

• Cloth: Acloth(92Ktriangles)hasahighnumberofselfcollisions(Figure7(d)).Ouralgorithmreducesthenumberof<br />

elementarytestsby 11.5x.<br />

• Flamenco:Thisbenchmark(49Ktriangles)hasmanyinter<strong>and</strong>intra-objectcollisions(Figure7(e)).TheCCDalgorithm<br />

withthede<strong>for</strong>mingnon-penetrationfilterreducesthenumber<br />

ofelementarytestsby 17x.<br />

Fig.8highlightsthecullingefficiencyofouralgorithmbycomparingthenumberofelementarytestsper<strong>for</strong>med.Asshowninthefigure,theelementarytestsaredramaticallyreducedby<br />

10 − 58.7x.<br />

Also,bypreventingthecomputationofthesefalsepositives,we<br />

achieve 1.5−3.5ximprovementonoverallper<strong>for</strong>manceincomparisonwiththeICCDalgorithm[Tangetal.2009a],<strong>and</strong><br />

2.4 − 8.2x<br />

improvementoverR-Triangles[Curtisetal.2008](Table.1). For<br />

BenchmarkFlamenco,itsaveragerunningtimeperframeinR-<br />

Triangles,ICCD,<strong>and</strong>oursystemareshowninFigure9.<br />

5 Analysis<strong>and</strong>Comparison<br />

Inthissection,weanalyzeourresults<strong>and</strong>compareagainstprior<br />

methods.<br />

5.1 Analysis<br />

Althoughtheper<strong>for</strong>manceofourde<strong>for</strong>mingnon-penetrationfilter<br />

varieswithdifferentbenchmarks,weareabletoobtainhighculling<br />

(a) Balls<br />

(b) Falling Alphabets<br />

(d) Cloth<br />

(c) Princess (e) Flamenco<br />

Figure7:Benchmarks:Allthebenchmarkshavemultiplesimulationsteps.Weper<strong>for</strong>mCCD,includingself-collisions,betweendiscretestepsofthesimulation<strong>and</strong>computethefirsttime-of-contact.<br />

efficiencyinmanyofthem.Thebasicfiltertestproposedinthenoncoplanartheoremsisconservative.<br />

Intuitively,ournon-coplanar<br />

theoremeliminatestheneedtoper<strong>for</strong>mexacttestsonpairswhere<br />

no‘penetrations’occuralongthecontinuoustrajectory. Inpractice,duetothetemporalcoherencebetweensubsequentframes,the<br />

’penetrations’onlyoccurinfrequently. Fig.10showsthechangingofratiosofprimitivepairsinwhichno‘penetration’occurs<br />

duringthetimeinterval. Inthefigure,theratiosstayabove 80%,<br />

<strong>and</strong>arechangingsmoothly.Duetocoherence,ourde<strong>for</strong>mingnonpenetrationfiltercanbequiteeffective.<br />

WeusetheInterval-Newtonmethodtosolvethecubicequations.<br />

Anelementarytesttakesaboutroughly155additions,217multiplications,<strong>and</strong>6divisionsonaverage(all<strong>for</strong>floatpointvalues).<br />

Ontheotherh<strong>and</strong>,thede<strong>for</strong>mingnon-penetrationfilteronlyneedto<br />

per<strong>for</strong>m29additions<strong>and</strong>40multiplicationsonaverage.Inpractice,<br />

thespeedofade<strong>for</strong>mingnon-penetrationfilterisabout5.5-10.2x<br />

fasterthananexacttest.<br />

5.2 Comparison<br />

Inthissection,wecompareournovelalgorithmwithpriorculling<br />

algorithms.<br />

Bounding volume based culling: Bounding volumes (e.g.<br />

spheres [Hubbard 1993; Palmer <strong>and</strong> Grimsdale 1995; Bradshaw<strong>and</strong>O’Sullivan2004],AABBs[v<strong>and</strong>enBergen1997],k-

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!