Upload
parson
View
39
Download
2
Tags:
Embed Size (px)
DESCRIPTION
Appearance-guided Synthesis of Element Arrangements by Example. Thomas Hurtut , Pierre- Edouard Landes , Joëlle Thollot Yann Gousseau Remy Drouilhet , Jean-Fran ç ois Coeurjolly. Motivation. Mucha , The Lady of the Camellias , 1896. Hokusai, Mount Fuji in Clear Weather , 1823. - PowerPoint PPT Presentation
Citation preview
Appearance-guided Synthesisof Element Arrangementsby Example
Thomas Hurtut, Pierre-Edouard Landes, Joëlle ThollotYann GousseauRemy Drouilhet, Jean-François Coeurjolly
MotivationMucha, The Lady of the Camellias, 1896
Klimt, Portrait of Adele Bloch-Bauer I, 1907
Hokusai, Mount Fuji in Clear Weather, 1823
Our Goals
Vector-based inputs
By-example approach
Controllable synthesis
Capture of non-uniform distributions
VectorRaster
Our method
Ijiri et al. ‘08 Barla et al. ’06 RRBarla et al. ’06 EGExtension to vector primitives
• Unconstrained positions• More than pixel colors
StatisticalmodelingGuo et al. ‘01
Statistics matchingPortilla et al. ‘01
Patch-basedEfros et al. ‘01Dischler et al. ’02Kwatra et al. ‘03
Pixel-basedEfros et al. ‘99Ashikhmin et al. ‘01Lefebvre et al. ‘06
Non-parametric Parametric
Example-based Texture Synthesis
Local neighborhood matching
Our method
Non-parametric Parametric
Example-based Texture Synthesis
Hexagonal distributionsNear-regular to
random distributions
Global analysis
Regularity
Pixel grid Delaunay triangulation
Ijiri et al. ‘08 Barla et al. ’06 RRBarla et al. ’06 EG
Our method
Our Method
• Statistical analysis of the element distribution
• Model the spatial interactions between elementsOur model = multi-type point processes
• But, many model parameters from limited input
Our Method
RealizationModel fitting
Untractable if continuous interactions considered
Our Method
Categorization RealizationModel fittingModel fitting Realization
Untractable if continuous interactions considered
Our Method
Categorization RealizationModel fitting
Simplification by Categorization
2 perceptual principles:• Law of similarity [Kohler ‘76]
Visually-similar elements perceived as a unit
• Texton discrimination [Julesz ‘86]Visually-dissimilar elements pre-attentively discriminated
Element Description
Area
Orientation
Elongation
# Extremities
# Crossings
Area
Orientation
Elongation
# Extremities
# Crossings
Descriptive features
Curve-based elements
Inspired from Julesz’ perceptual studies
Feature Analysis
Reduction to 1D
AreaArea Orientation Elongation #Extremities #CrossingsOrientation Elongation #Extremities #Crossings
2-fold analysis
Categorization by a contrario mode-seeking
Categorization Results
Categorization Results
Categorization Results
Our Method
RealizationModel fitting
Categorization
Our Method
RealizationModel fitting
Categorization
Modeling the Spatial Interactions
How often? At what distance?
Modeling the Spatial Interactions
At what distance?How often?
Interactions Between Categories
Strauss hard-core interaction model
Let us consider the term
distance
Parameter Estimation (details in paper)
• Pseudo-likelihood maximization
• Minimal pair-wise Euclidean distance
• Analysis of the Ripley function
How often?
At what distance?
Our Method
Realization
Categorization Model fitting
Our Method
Realization
Categorization Model fitting
Arrangements as Model Samples
Monte-Carlo Markov Chain sampling
Iterative procedure: random elementary perturbations
timesteps
Output
Arrangements as Model Samples
timesteps
Output
Birth Death
Monte-Carlo Markov Chain sampling
Iterative procedure: random elementary perturbations
Arrangements as Model Samples
timesteps
Birth Death
Output
Monte-Carlo Markov Chain sampling
Iterative procedure: random elementary perturbations
Arrangements as Model Samples
timesteps
Output
Monte-Carlo Markov Chain sampling
Iterative procedure: random elementary perturbations
Our Method
Categorization Model fitting Realization
Output
Output
Output
Output
Comparison with Related Work
Barla et al. ‘06 Ours
OursBarla et al. ‘06
Comparison with Related Work
Comparison with Related Work
OursBarla et al. ‘06
Element Density Control
Element Density Control
Barla et al. ‘06OursMucha, The Lady of the Camellias, 1896
« Emulating the Masters »
« Emulating the Masters »
Hokusai, Mount Fuji in Clear Weather, 1823
Over-categorization
Output
Handling of Regularity
Output
Interactions Between Centroids
Output
Future Work & Conclusions
• Possible improvements• Higher-order interactions for regularity• Better representation for anisotropic elements• Quality-driven stopping criterion for faster synthesis
• Contributions• Global input analysis• Appearance-driven synthesis• Wider range of distributions supported
Behind the Scenes (Parameter Estimation)
Model to fit
Full arrangement
Estimation ofby log-pseudo-likelihood maximization
IntuitionStatistical “explanation” of the element positions
Behind the Scenes (Parameter Estimation)
Model to fit
Full arrangement
Estimation of by Ripley function minimization
IntuitionEvaluate distance where element distribution gets close to random
If distribution purely random
If distribution more regular