Title: Importance Sampling from Product of the BRDF and the Illumination using Spherical Radial Basis Funct
1Importance Sampling from Product of the BRDF and
the Illumination using Spherical Radial Basis
Functions
- Student Qing-Zhen Jiang
- Advisor Prof. Zen-Chung Shih
- Institute of Multimedia EngineeringNational
Chiao Tung University
2Outline
- Introduction
- Related Work
- Background of SRBFs
- System Overview
- Off-Line SRBF Fitting Process
- Run-Time Rendering Process
- Results
- Conclusions and Future Works
3Introduction
4Introduction
- The goal of global illumination is to solve the
rendering equation Kajiya 1986 - Monte Carlo Technique
- To increase efficiency, importance sampling is a
powerful technique.
5Introduction
Uniform Sampling
6Introduction
Importance Sampling
7Introduction
X
Environment map importance sampling
BRDF importance sampling
From Clarberg et al. 2005
8Complex Models
measured BRDF data
HDR environment map
Matusik et al. 2003
St. Peters Basilica
9Related Work
10Related Work BRDF Importance Sampling
- Lafortune Model (Multiple cosine lobes)
- Lafortune et al. SIGGRAPH 1997
- Non-linear fitting
- Matrix factorization
- Lawrence et al. SIGGRAPH 2004
- Non-negative Matrix Factorization (NMF)
- Wavelet
- Matusik et al. SIGGRAPH 2003
- Lalonde PhD thesis 1997
- Spherical Radial Basis Functions (SRBFs)
- Weng and Shih Master thesis 2006
- Non-uniform and non-negative fitting
11Related Work Environment Map Importance
Sampling
- According the energy distribution
- Cohen and Debevec 2001, Agarwal et al. 2003,
Kollig and Keller 2003, Ostromoukhov 2004. - Based on clustering algorithm or hierarchical
tiling scheme - Spherical Harmonics
- Ramamoorthi and Hanrahan 2002
- Wavelet
- Ng R. et al. 2003
12Related Work Sampling from Product
Distributions
- Bidirectional sampling
- Bruke et al. Eurographics 2005
- Rejection sampling
- Sampling-importance resampling
13Related Work Sampling from Product
Distributions
- Wavelet Important Sampling
- Clarberg et al. SIGGRAPH 2005
14Related Work Sampling from Product
Distributions
- Wavelet Important Sampling
- Clarberg et al. SIGGRAPH 2005
15Background of SRBFs
16Background of SRBFs
Center
Bandwidth
Coefficient
Legendre Polynomial
?
?
?
17Background of SRBFs
- Spherical singular integral
-
- Gaussian SRBF
-
- There is a simple mathematical for the
convolution of two Gaussian SRBF kernels -
-
- where
18System Overview
19System Overview Off-Line SRBF Fitting Process
Lighting Direction
Viewing Direction
Measured BRDF Data
20System Overview Off-Line SRBF Fitting Process
Environment map
21System Overview Run-Time Rendering Process
Environment map
BRDF
Viewing Direction
Product
Generate Samples from each SRBF
22Off-Line SRBF Fitting Process
23Non-uniform and Non-negative SRBF Fitting
Algorithm HDR environment map
Initial Guess
optimize centers
optimize bandwidths
L-BRFS-B solver
No
optimize coefficients
SE lt t Iters gt n
Yes
terminate
24Initial Guess
Accept!!
Reject!!
Coverage-Weighted Square of Intensity
25Non-uniform and Non-negative SRBF Fitting
Algorithm HDR environment map
St. Peters Basilica
Reconstructed from 300 SRBFs
26Non-uniform and Non-negative SRBF Fitting
Algorithm HDR environment map
Uffizi Gallery
Reconstructed from 300 SRBFs
27Non-uniform and Non-negative SRBF Fitting
Algorithm Measured BRDF Data
-
- Weng and Shih Master thesis 2006
28Run-Time Rendering Process
29Product of BRDF and illumination
- Product of Gaussian SRBF
-
- where , ,
- We prune smaller efficient terms to reduce
computation cost.
30Importance sampling
- Monte Carlo estimator
- Choose a density function p that is similar to
the integrand f -
- Multiple Importance Sampling
- Veach and Guibas SIGGRAPH 95
??
31Multiple Importance Sampling
- Combining estimators
- Weighted-average of all estimators
- Weights depend on the sampling positions
-
-
Product of BRDF and illumination
PDF computed from SRBF
??
Weight
32Allocate Samples for each SRBF
- The density of each SRBF kernel can be estimated
by its integral - Integral can be calculated easily with spherical
singular integral property. - The number of samples that should be taken from
each SRBF is determined by -
-
- Calculate a 1D CDF for these probabilities.
- Allocating samples by generating uniform
variables in 0, 1
33Generate Samples for each SRBF
- Choosing azimuth angle f
- Uniform sample in 0, 2p
- Choosing elevation angle ?
- Metropolis Random Walk Algorithm
- Use random mutations to produce a set of samples
with a desired density
34Results
35Results
10 samples
40 samples
60 samples
Environment map Grace cathedral
100 samples
100 Samples (BRDF Only)
80 samples
36Results
Environment mapSt. Peters Basilica
60 samples
37Conclusions and Future Works
38Conclusions and Future Works
- Conclusions
- A new method for sampling the products of complex
functions. - SRBF importance sampling can render noise-free
images using 60-100 samples per pixel. - Future Works
- The major computation cost for ray tracing is the
visibility testing. - Generate samples smarter based on some heuristic
approach.