Image-Guided Weathering: A New Approach Applied to Flow Phenomena
description
Transcript of Image-Guided Weathering: A New Approach Applied to Flow Phenomena
Image-Guided Weathering:Image-Guided Weathering:A New Approach Applied to Flow A New Approach Applied to Flow
PhenomenaPhenomenaC. BoschC. Bosch11, P. Y. Laffont, H. Rushmeier, , P. Y. Laffont, H. Rushmeier,
J. Dorsey, G. DrettakisJ. Dorsey, G. Drettakis
Yale University – REVES/INRIA Sophia AntipolisYale University – REVES/INRIA Sophia Antipolis1 1 Currently at ViRVIG, University of GironaCurrently at ViRVIG, University of Girona
Aging and WeatheringAging and Weathering
Essential for modeling urban environments Governed by physical, chemical and biological processes
Flow effectsFlow effects
Particularly complex Flow over the scene (global effect)
Material properties (local effect)
Aging and Weathering in CGAging and Weathering in CG
Physically-based simulation Difficult to get the desired effect
Texture synthesis Restricted by input information
Global effects particularly hard
MotivationMotivation
Physically-based simulation More flexible, allows global effects
Two main difficulties Choosing appropriate parameters to achieve a given effect
Obtaining realistic visual detail
Image-Guided WeatheringImage-Guided Weathering
Use images to guide simulation Flow stains as a representative case
Exemplar
New simulation
Overview (I)Overview (I)
Extract data from exemplars Color information
Simulation parameters
High frequency details
Si = 1.301 rt = 0.252kS = 0.0201 at = 0.404kD = 0.0807 T = 803ka,t = 0.021
Exemplar
Data
Overview (II)Overview (II)
Simulate new effects on scenes
Si = 1.301 rt = 0.252kS = 0.0201 at = 0.404kD = 0.0807 T = 803ka,t = 0.021
Data
Related WorkRelated Work
Simulation Phenomenon-specific [Merillou08]
Flow stains [Dorsey96; Chen05; Endo10]
Capture-and-transfer (synthesis) Single image [Wang06; Xue08]
Acquisition systems [Gu06; Mertens06; Sun07; Lu07]
Inverse procedural textures [Bourque04; Lefebvre00]
Flow modelFlow model
Particle-based simulation [Dorsey96] Absorption, solubility and deposition
Stain concentration maps
Parameters Particles: mass (m), Si
Stain material: kS, kD
Target materials: a, ka, roughness (r)
Simulation: time (t), particle rate (N)
Extracting StainsExtracting Stains
Based on Appearance Manifolds [Wang06]
ExemplarAppearance Manifold
Degree Map
Degree map = Stain concentration map
Error Simulation
Parameter FittingParameter FittingInput stain Degree map
Si = 1kS = 0.04kD = 0.04rt = 0.2at = 0.3ka,t = 0.05T = 300
Si = 1kS = 0.04kD = 0.04rt = 0.2at = 0.3ka,t = 0.05T = 300
Initialparameters
New parameters
Si = 1.3kS = 0.02kD = 0.08rt = 0.25at = 0.4ka,t = 0.02T = 803
Si = 1.3kS = 0.02kD = 0.08rt = 0.25at = 0.4ka,t = 0.02T = 803
Error < threshold or max. iterations
StopStop
Proxy geometry
target
(Levenberg-Marquardt) [Lourakis04]
image plane
source
Improving FittingImproving Fitting
1. Stain distribution along the source Accumulate degree from bottom to top
Improving Fitting (II)Improving Fitting (II)
2. Flow deflection along the target Compute local degree distribution (~vector field)
Error Simulation
Input stain Degree map
Si = 1kS = 0.04kD = 0.04rt = 0.2at = 0.3ka,t = 0.05T = 300
Si = 1kS = 0.04kD = 0.04rt = 0.2at = 0.3ka,t = 0.05T = 300
Initialparameters
New parameters
Si = 1.3kS = 0.02kD = 0.08rt = 0.25at = 0.4ka,t = 0.02T = 803
Si = 1.3kS = 0.02kD = 0.08rt = 0.25at = 0.4ka,t = 0.02T = 803
Error < threshold or max. iterations
StopStop
Proxy geometry
target
(Levenberg-Marquardt) [Lourakis04]
image plane
source
Parameter Fitting (II)Parameter Fitting (II)
Vector field
Stain distribution
Fitting Results (w/o vector field)Fitting Results (w/o vector field)
Exemplar Degree Map SimulationUsing source distribution
Fitting Results (w/o vector field)Fitting Results (w/o vector field)
Exemplar Degree Map Simulation
Fitting Results (w/ vector field)Fitting Results (w/ vector field)
Exemplar
Degree Map Simulation w/o vfield
Fitting Results (w/ vector field)Fitting Results (w/ vector field)Exemplar Degree Map Simulation
Fitting Results (w/ vector field)Fitting Results (w/ vector field)Exemplar Degree Map Simulation
Fitting Results (Complex Targets)Fitting Results (Complex Targets)
Exemplar Degree Map Simulation
Stain DetailStain Detail
Simulation lacks spatial variations (high-frequency detail)
Degree Map Simulation
Exemplar
Detail MapsDetail Maps
Extract detail by image difference Use guided texture synthesis [Lefebvre05] Detail maps will modify stain adhesion
Degree Map Simulation Difference
Detail Map
Simulating New StainsSimulating New Stains
Link data to stain sources and targets Parameters, detail maps, color
Use 1D texture synthesis for distributions
Run flow simulation Flow deflected by target geometry (+ disp. map)
Color TransferColor Transfer
Transfer stain color from input image Background mixed with stain everywhere
Non-linear relationship between color and degree
Use per-pixel warping
background color
target background
fully stained
ResultsResults
Results (II)Results (II)
Results (III)Results (III)
Results (IV)Results (IV)
PerformancePerformance
Preprocessing Degree map: 1-3 minutes
Fitting: 30-60 minutes (500 iter., ~256x512)
Detail synthesis: 1-2 minutes (1024x1024)
Final simulation Stain simulation: 2-5 minutes/stain
Color warping: 5-8 seconds/stain (1024x1024)
LimitationsLimitations
Good extraction from background Fitting: Not true physical estimations Detail maps: Depend on appropriate fit Computation time
ConclusionsConclusions
New approach to acquire simulation data from photographs Solves parameter estimation from images
Combines simulation with data-driven methods Appearance manifold, texture synthesis, …
Fills the gap between data-driven and simulation
Easy to use
Natural variations (including global effects)
Future workFuture work
Extend to other weathering phenomena Deal with large scale scenes
Fast simulation, global effects, …
AcknowledgementsAcknowledgements
Visiting grant U.Girona ANR project (ANR-06-MDCA-004-01) ERCIM “Alain Bensoussan” Fellowship Autodesk (Maya/MentalRay) Coding help: Li-Ying, Su Xue Scene treatment: S. Close and F. Andrade-Cabral
Thank youThank you