CS348B Lecture 9Pat Hanrahan, Spring 2005 Overview Earlier lecture Statistical sampling and Monte...
-
Upload
tamsyn-dickerson -
Category
Documents
-
view
215 -
download
0
Transcript of CS348B Lecture 9Pat Hanrahan, Spring 2005 Overview Earlier lecture Statistical sampling and Monte...
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Overview
Earlier lecture Statistical sampling and Monte Carlo integration
Last lecture Signal processing view of sampling
Today Variance reduction Importance sampling Stratified sampling Multidimensional sampling patterns Discrepancy and Quasi-Monte Carlo
Latter Path tracing for interreflection Density estimation
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Cameras
( , , ) ( ) ( ) cosT A
R L x t P x S t dA d dt
Source: Cook, Porter, Carpenter, 1984Source: Mitchell, 1991
Depth of FieldMotion Blur
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Variance
1 shadow ray per eye ray 16 shadow rays per eye ray
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Variance
Definition
Variance decreases with sample size
2
2 2
[ ] [( [ ]) ]
[ ] [ ]
V Y E Y E Y
E Y E Y
2 21 1
1 1 1 1[ ] [ ] [ ] [ ]
N N
i ii i
V Y V Y NV Y V YN N N N
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Variance Reduction
Efficiency measure
If one technique has twice the variance as another technique, then it takes twice as many samples to achieve the same variance
If one technique has twice the cost of another technique with the same variance, then it takes twice as much time to achieve the same variance
Techniques to increase efficiency
Importance sampling
Stratified sampling
1Efficiency
Variance Cost
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Biasing
Previously used a uniform probability distribution
Can use another probability distribution
But must change the estimator
~ ( )iX p x
( )
( )i
ii
f XY
p X
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Unbiased Estimate
Probability
Estimator
~ ( )iX p x
( )
( )i
ii
f XY
p X
( )[ ]
( )
( )( )
( )
( )
ii
i
i
i
f XE Y E
p X
f Xp x dx
p X
f x dx
I
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Importance Sampling
( )( )
[ ]
1( )
[ ]
1
f xp x dx dx
E f
f x dxE f
( )( )
[ ]
f xp x
E f
Sample according to f
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Importance Sampling
Variance
2 2[ ] [ ] [ ]V f E f E f 2
2
2
2
( )[ ] ( )
( )
( ) ( )
( ) / [ ] [ ]
[ ] ( )
[ ]
f xE f p x dx
p x
f x f xdx
f x E f E f
E f f x dx
E f
( )( )
[ ]
f xp x
E f
( )( )
( )
f xf x
p x
Sample according to f
2[ ] 0V f
Zero variance!
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Example
method Sampling
function
variance Samples needed for standard error of 0.008
importance
(6-x)/16 56.8N-1 887,500
importance
1/4 21.3N-1 332,812
importance
(x+2)/16 6.4N-1 98,432
importance
x/8 0 1
stratified 1/4 21.3N-3 70
4
0
8xdxI Peter Shirley – Realistic Ray Tracing
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Examples
Projected solid angle
4 eye rays per pixel100 shadow rays
Area
4 eye rays per pixel100 shadow rays
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Irradiance
Generate cosine weighted distribution
2
( )cosi i i i
H
E L d
( ) cosp d d
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Cosine Weighted Distribution
2
0
22
0
2
0
))(cos21(2)sin()cos()cos(
2
dddH
)()()sin()cos(
),()sin()cos(
),(
pppddddp
2
1)( p
21 U
12 U
22
1)( 0
0
0
0
dP
)sin()cos(2)( p
)(sin)(sin)sin()cos(2)( 02
00
20
0
0
dP
)(sin 22 U )arcsin( 2U
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Sampling a Circle
1
2
2 U
r U
Equi-Areal
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Shirley’s Mapping
1
2
14
r U
U
U
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Stratified Sampling
Stratified sampling is like jittered sampling
Allocate samples per region
New variance
Thus, if the variance in regions is less than the overall variance, there will be a reduction in resulting variance
For example: An edge through a pixel
21
1[ ] [ ]
N
N ii
V F V FN
2 1.51
1 [ ][ ] [ ]
NE
N ji
V FV F V F
N N
1
1 N
N ii
F FN
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Mitchell 91
Uniform random Spectrally optimized
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Discrepancy
x
y
( , )( , )
( , ) number of samples in
n x yx y xy
N
A xy
n x y A
,max ( , )Nx y
D x y
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Theorem on Total Variation
Theorem:
Proof: Integrate by parts
1
1( ) ( ) ( )
N
i Ni
f X f x dx V f DN
( ) ( )1ix x x
x N
1
0
( )( ) [ 1]
( )( )
( ) ( )( ) ( )
( )( )
i
N N
x xf x dx
Nx
f x dxxf x f x
f x dx x dxx x
f xD dx V f D
x
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Quasi-Monte Carlo Patterns
Radical inverse (digit reverse)
of integer i in integer base b
Hammersley points
Halton points (sequential)
2 1 0
0 1 2( ) 0.i
b i
i d d d d
i d d d d
1 1 .1 1/22 10 .01 1/4
3 11 .11 3/44 100 .001 3/85 101 .101 5/8
2( )i
2 3 5( / , ( ), ( ), ( ), )i N i i i 1log
( )d
N
ND O
N
2 3 5( ( ), ( ), ( ), )i i i log
( )d
N
ND O
N
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Hammersly Points
2 3 5( / , ( ), ( ), ( ), )i N i i i
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Edge Discrepancy
ax by c
Note: SGI IR Multisampling extension: 8x8 subpixel grid; 1,2,4,8 samples
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Low-Discrepancy Patterns
Process 16 points 256 points 1600 points
Zaremba 0.0504 0.00478 0.00111
Jittered 0.0538 0.00595 0.00146
Poisson-Disk
0.0613 0.00767 0.00241
N-Rooks 0.0637 0.0123 0.00488
Random 0.0924 0.0224 0.00866
Discrepancy of random edges, From Mitchell (1992)
Random sampling converges as N-1/2
Zaremba converges faster and has lower discrepancy
Zaremba has a relatively poor blue noise spectraJittered and Poisson-Disk recommended
CS348B Lecture 9 Pat Hanrahan, Spring 2005
High-dimensional Sampling
Numerical quadrature
For a given error …
Random sampling
For a given variance …
1
1 1~
dE
n N
1/ 21/ 2
1~ ~E V
N
Monte Carlo requires fewer samplesfor the same error in high dimensional spaces
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Block Design
a
bc
d
a
bc
d
a
b
c
d
a
b
c
d
Alphabet of size n
Each symbol appears exactly once ineach row and column
Rows and columns are stratified
Latin Square
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Block Design
a
a
a
a
N-Rook Pattern
Incomplete block design
Replaced n2 samples with n samples
Permutations:
Generalizations: N-queens, 2D projection
1 2( ( ), ( ), ( ))di i i
( {1,2,3,4}, {4,2,3,1})x y
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Space-time Patterns
Distribute samples in time
Complete in space
Samples in space should have blue-noise spectrum
Incomplete in time
Decorrelate space and time
Nearby samples in space should differ greatly in time
15
4
10
1
3
13
6
8
0
11
5
14
12
2
9
7
Pan-diagonal Magic Square
15
4
10
1
3
136
8
0 11
5
14 12
2
9
7
Cook Pattern
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Path Tracing
4 eye rays per pixel16 shadow rays per eye ray
64 eye rays per pixel1 shadow ray per eye ray
Complete Incomplete
CS348B Lecture 9 Pat Hanrahan, Spring 2005
Views of Integration
1. Signal processing
Sampling and reconstruction, aliasing and antialiasing
Blue noise good
2. Statistical sampling (Monte Carlo)
Sampling like polling
Variance
High dimensional sampling: 1/N1/2
3. Quasi Monte Carlo
Discrepancy
Asymptotic efficiency in high dimensions
4. Numerical
Quadrature/Integration rules
Smooth functions