qyu defense

8/12/2019 Qyu Defense

1/134

a Ph.D. Defense by

Qizhi YuUnder the Advisements ofDr. Fabrice NeyretDr. Eric Bruneton

November 17, 2008

Grenoble Institute of Technology (INPG)

1


2/134

Introduction

Previous work

Strategy overview

Contributions

Conclusion

2


3/134


4/134

RiversHydrology

Hydraulics Geomorphology

Ecology

4


5/134

RiversHydraulics

Hydrology

Computergraphics

Geomorphology

Ecology

5


6/134

ObjectiveSynthesize visually convincing rivers

Study contentModeling River shape & surface details

Animating Water motion in rivers.

6


7/134

Many applications, need more studies

EA: CrysisGoogle Earth

7


8/134

Multi-scaleGeometry

Kilometer-scale lengthmillimeter-scale waves

Water motion Kilometer-scale mean flow

millimeter-scale fluctuation

Complicated physicsTurbulence Surface phenomena

8


9/134

Modeling and animating rivers

ConstraintsReal-timeScalabilityControllability

Realism

25 fps or more

9


10/134



Realism

Very long or unbounded riversCamera moves arbitrarily

10


11/134



Realism Intuitive handles for controllingappearance and behavior of rivers

11


12/134



RealismAnimated surface detailswith temporal and spatial continuity

12


13/134



Realism

13


14/134

Introduction

Previous work

Strategy overview

Contributions

Conclusion

14


15/134

3D Navier-Stokes simulation

2D depth-averaged simulation

2D simulation

Surface wave models (2D)

15


16/134



2D simulation


16


17/134

Incompressible Navier-Stokes equationsMomentum conservation

Volume conservation

Boundary conditions

Computational Fluid Dynamics (CFD)Numerical methods

17


18/134

CFD CG fluid animationStable solver [stam99]

Two approachesEulerian : defines quantities at fixed pointLagrangian: defines quantities at particles

18


19/134

Water animation [EMF02]Solve NSE numerically on a grid to get velocities

Use level-set to track water-air interface

19


20/134

Pouring water in a glass

[EMF02]15 minutes per frame

55 x 120 x55 grids

Computationally expensive!

20


21/134

Poorly scalableCG (stable solver): O(N^3)

Difficult to control for artistsWater Behavior Initial values, boundary conditionsNo intuitive relation

21


22/134

Smoothed Particle Hydrodynamics (SPH) [MCG03]Solve NSE in the Lagrangian formalism

Compared with Eulerian approachEasier adaptive to complex domainDifficult to reconstructa smooth surface

For our purposeSimilar problemsas Eulerian approach

22

2200 particles5 fps


23/134



2D simulation


23


24/134

2D Shallow Water model [Mol95]Commonly used in Hydraulics for simulating riversAssumptions Hydrostatic approximation

No vertical water motionIntegrate the NS equations along vertical directionUnknowns:

depth-averaged velocity & elevation of water surface

24


25/134

PropertiesA lot faster than 3D N-S simulation

Loss some 3D surface features (e.g. overturning )Shallow waves ( wavelength >> depth)

For our purposeStill too expensive, especially for large riversBounded domain (like other simulation).

25


26/134

Linear wave equation [KM90]Simplified from shallow water model

Assumptions constant water depth, no advection term

Properties Fast, cant simulate river flow

26


27/134

[Irving et al. 06]

20 processors25 minutes per frame

27


28/134



2D simulation


28


29/134

Simulate 2D velocity by solving 2D N-Sno surface elevation simulated

Use tricks for surface elevationPressure [CdVL95]Noise [TG01]

29


30/134



2D simulation


30


31/134

AssumptionDeep water: wave length


32/134

AssumptionHeight field

A procedural methodParticles on surfaces, advected with a fixed speedEach carries a wave shape functionSuperpose all particles height field

PropertiesImitate object-water interactionNo water flow

32


33/134

Superpose sine waves [FR86, Pea86]Dynamic wave tracing [GS00]Ship wave [Gla02]

33

[Gla02] [GS00] [Pea86]

Not for river flow


34/134

Many work on water or wave animation (CG),river simulation (Hydraulics)

None for river animation under ourconstrains:

Real-time

ScalabilityControllability

34


35/134

Introduction

Previous work

Strategy overview

Contributions

Conclusion

35


36/134

Model river aspects in three scales, fromcoarse to fine

36


37/134


38/134

Model river aspects in three scalesMacro-scale: river shape & mean water surface

Meso-scale: individual & structured waves

38


39/134

Model river aspects in three scalesMacro-scale: river shape & mean water surface

Meso-scale: individual & structured wavesMicro-scale: continuous field of small waves

39


40/134

We need river velocityCause of many meso-scale phenomena

Advect surface featuresModel water motion in three scales

Macro-scale: mean flow

Meso-scale: individual perturbationsMicro-scale: continuous irregular fluctuations

40


41/134

We need river velocityCause of many meso-scale phenomena

Advect surface featuresModel water motion in three scales

Macro-scale: mean flow

Meso-scale: individual perturbationsMicro-scale: continuous irregular fluctuations

We wont solve ALL phenomena in this thesis . 41


42/134

Introduction

Previous work

Strategy overview

Contributions1: Macro-scale2: Meso-scale3: Micro-scale

Conclusion42


43/134

Introduction

Previous work

Strategy overview

Contributions1: Macro-scale

2: Meso-scale3: Micro-scale

Conclusion43


44/134

GoalShape of rivers

Mean flow of rivers

44


45/134

GoalShape of rivers

Mean flow of rivers

GIS or previous work[KMM88]

45


46/134

Inputriver shape (described as a network)

46


47/134

Assumption a 2D steady flow

Visually convincing velocityDivergence free IncompressibleBoundary-conformingFlowing from source to sink (given flow rate Q)

Continuous

Requirements of algorithmsFast, scalable and controllable

47


48/134

Some existing work [BHN07] suggestusing stream function to get divergence-freevector field

48


49/134

Stream function is defined such that

Incompressibility

49


50/134

Properties of stream functionConst along boundariesRelates to the volume flow rate

Extend to a river networkGiven flow rates and a river network all boundary values

50


51/134

Given flow rates, and boundary valuesHow to determine the internal field ?

51


52/134

AssumptionIrrotational (potential) flow

52


53/134

Observe a numerical solution of a Laplaceequation

Streamlines (isocurve of stream function) 53


54/134

[GW78]Interpolant of the Inverse-Distance Weightedinterpolation (IDW) [She68] similar to theharmonic functions.

We adapt IDW

local for the performance reasonsprovide parameters for controlling velocity profile

54


55/134

55

d distance to boundariesf smooth functions search radius

p parameters


56/134

Our result Numerical solution

56


57/134

Finite difference

57


58/134

Interpolation relies heavily on distance queryAcceleration needed

Combine with tile-based terrain [BN07]Generate an acceleration data structure in eachnewly created terrain on-the-fly

Please see the thesis for more details .

58


59/134

59


60/134

Procedural river flowFast

Scalable Calculate at needed Velocity locally dependent

Controllable Control velocity: flow rates, interpolation parameters Edit shape of river on-the-fly

60


61/134

Introduction

Previous work

Strategy overviewContributions

1: Macro-scale


Conclusion61


62/134

GoalModeling individual & structured wave features onriver surfaces, with our constraints. Real-time Scalability Controllability

Quality

62


63/134

63Real scene


64/134

High-resolution required for simulation andrendering

64


65/134

Construct the vector features from a givenvelocity field without numerical simulation

Shockwave

Ripples

65


66/134

ProblemsNeed to be improved robustness & efficiency

No solution for surface reconstruction and rendering

66


67/134

Improve on existing model [NP01]Result Mean flow shockwave curves (wave crests) Animated by adding perturbation to the mean flow

67


68/134

Improve on existing model [NP01]Result Mean flow shockwave curves (wave crests) Animated by adding perturbation to the mean flow

Macro-scale

Meso-scale, [WH91]

68


69/134

Improve on existing model [NP01]Result Mean flow shockwave curves (wave crests) Animated by adding perturbation to the mean flow Very efficient

69


70/134

Improve on existing model [NP01]

Construct appropriate representation from

wave features for high-quality rendering

70


71/134

lo-res base water surface + hi-res wave surface

71


72/134


Macro-scale

72


73/134


Meso-scale

73


74/134

Feature-aligned mesh reducesgeometric aliasing ( normal-noise)

Not feature-aligned Feature-aligned74


75/134

Define wave surface as sweeping a waveprofile along the wave curve

User defined Wave profile

Wave curve

75

Water surfacemesh


76/134

Sample by a quad meshaligned the wave curve

v 76

Wave curve


77/134

Accurate normals from the wave profile

u

v

B

NT

P(u,v)

v 77


78/134

Mesh stitching ?Re-mesh base surface at each frame, tooexpensive

We solve it in the rendering stage

Please refer to the thesis for more details

78


79/134

79


80/134

Simply draw two wave strips with Z-buffer

80


81/134

Generate a dedicated mesh at crossing

81


82/134

Final result

82


83/134

83


84/134

Approach: feature-based vector simulationSimulation: construct & animate vector featuresRendering: featured-based representation

84


85/134

Introduction

Previous work

Strategy overviewContributions

1: Macro-scale


Conclusion

85


86/134

Goal Modeling small scale animated surface features

Approachdynamic textures

Two workWave sprites Focus on performanceLagrangian texture advection Focus on quality

86


87/134

IntroductionPrevious workStrategy overviewContributions

1. Macro-scale2. Meso-scale

3. Micro-scale Wave sprites Lagrangian texture advection

Conclusion87


88/134

Sprite: a small textured element Sprites in texture world [LN03,LHN05]

to get large high-resolution texture , low memory Idea: combine animation + texture sprites

to get very large river with animated details,efficiently.

88


89/134

How should sprites behave for our purposes ?Sprites -> represent waves reconstructed texture should conserve the spectrum

Well distributed, avoiding holes and overcrowding The more overlapping, the more texture spectrum biasing

The density of sprites should be adaptive

Convey the flow motion

89


90/134

Dynamic adaptive samplingA set of particles in world space advected by flowKeep Poisson-disk distribution in screen space.

Attach a textured sprite to each particle

90


91/134

Dynamic adaptive samplingA set of particles in world space advected by flowKeep Poisson-disk distribution in screen space.

Attach a textured sprite to each particleWhy ?

91


92/134

Uniform densityOverlapping as little as possible

Easy to ensure spatial continuitySuperimposing sprites (with r=d) ensures no-holes

r r = ddiameter of poisson-disk

92


93/134

Dynamic adaptive samplingA set of particles in world space advected by flowKeep Poisson-disk distribution in screen space

Attach a sprite to each particle

Auto-adapt to distance

93


94/134

AlgorithmAdvect particles with the flow in world spaceDelete particles out of the view frustumDelete particles violating the minimum distancerequired by the Poisson-disk distribution (inscreen space)

Insert particles to keep Poisson-disk distribution

94


95/134

AlgorithmAdvect paticles with the flowDelete particles out of the view frustumDelete particles violating the minimum distancerequired by the Poisson-disk distribution (inscreen space)Insert particles to keep Poisson-disk distribution

Boundary-sampling algorithm [DH06]

95


96/134

Spatial continuitySmooth kernelConstrained sampling issues near boundary

Temporal continuityFading in/out

Please refer the thesis.

96


97/134

A set of sprites well distributedEach sprite Live in texture space maps to a portion of a reference texture

Reconstruct the global textureSprite has circular kernel in screen space , butellipse in object space So we superimpose them in screen space

97


98/134

98


99/134

99

Efficient GPU. Inspired from [LN05]


100/134

100


101/134

Wave-spritesTexture flow surface with scene-independentperformance (in real-time)

Limitation

No sprite deformation consideredSliding of texture between sprites bad especially in place where velocity gradient is high

101


102/134

IntroductionPrevious workStrategy overviewContributions

1. Macro-scale2. Meso-scale

3. Micro-scale Wave sprites Lagrangian texture advection

Conclusion102


103/134

A technique of dynamic textureConform to the input flowConserve texture properties (e.g. spectrum)

PurposeAugment coarse simulation with small scale

appearance

103


104/134

Advect texture coordinatesTexture follow flow and deformBut, over stretching destroy texture properties

Regenerate a textureAfter a delay: latency

Blend two de-phased textures Illusion of advection

104


105/134

How to choose a reasonable latency ? high bad conservation of spectrum low bad conformation to flowGood one: adapt to local flow condition

(deformation)

In [MB95], only one global value

105


106/134

Idea: adaptive local latency

Local deformation metrics

MIPmap -like approach:Multiple layers of texturesEach layer = Eulerian advection methodAssign different latency to each layer For each pixel, interpolate two nearest layersaccording to local s

106


107/134

latency of all layers are bounded in arange

e.g. For zero-velocity , the ideal latency should beinfinity close to still area, we cant choose agood latency value

Interpolation : not accurate

Eulerian formalismnot optimal in large sparse domain (clouds, fire)

107


108/134


109/134

Advected by flow

Dynamic Poisson-disk

distributiond

109


110/134


111/134

Init: regular grid

Kernel radius = d

Patch size > 2d

Map to a random portionStore (u, v) at nodes

U

V

111


112/134

Nodes advected by flow

112


113/134


Delete a patchExceed some deformationmetric

113


114/134


Delete a patchExceed some deformationmetric

Patch boundary intersectswith kernel

A new patch would be generated nearby automatically

by Poisson-disk distribution mechanism 114


115/134

Temporal & spatialInsert / delete temporalSmooth kernel spatial

Define various temporal and spatial weights on grid nodes

Please see details in the thesis

115


116/134

Encode all patches into one texture T patchTexcoords (u, v)Weights w(x, t)

Accessing the advected textureFor each pixel

Determine the patches covering current pixel Access reference texture via T patch Blending with weights (only kernel parts!)

116


117/134

117


118/134

Compare against Eulerian advectionFFT To evaluate the appearant spectrum

Optical flow To evaluate the appearant motion

Input reference texture 3-octave Perlin noise

118


119/134

Various input flow

Please see my webpage for more video results

119

Rotation Shear Free Boundary


120/134

120


121/134

121


122/134

We target textures specified by globalproperties, e.g. spectrum

Useful for natural flow

For non-noise texturesMany of them work wellHigh-structured ones Suffer from ghosting effects Future work: choose best match portion from reference

texture

122


123/134

123


124/134

124


125/134

A new texture advection methodLagrangian formalism Brings decorrelation of texture mapping and

regeneration eventsLocal patches Ensure continuous texture animation Provide accurate distortion metric

126


126/134

Introduction

Previous work

Strategy overview

Contributions

Conclusion

127


127/134

By using our modelsOne can achieve real-time , scalable , andcontrollable river animation with temporally andspatially continuous details on current desktop

128


128/134

Macro-scaleVelocity: more studies on parametersInfluence of slope of river bed

129


129/134

Meso-scaleHydraulic jumps, ship waves and wakes ...

130


130/134

Micro-scale I: wave spritesVarious reference textures: domain wise controlSprites density: adaptive to flow condition

131


131/134

Micro-scale II: Lagrangian texture advectionExtend to 3D volumeImprove: for high-structured texture

132[RNGF03]


132/134

Put models together

Integrate with existing systems

Google Earth, Proland [BN08], video games

133EA: CrysisGoogle Earth Proland


133/134

134

Grenoble Institute of Technology


134/134

a Ph.D. Defense by

Qizhi Yu

Under the Advisements of

Dr. Fabrice Neyret

Dr. Eric Bruneton

November 17, 2008

qyu defense

Documents