SAMPLING MEASURES & ERROR FORMULATIONS Cengiz ztireli Research - - PowerPoint PPT Presentation

Analysis of Sample Correlations for Monte Carlo Rendering

SAMPLING MEASURES & ERROR FORMULATIONS

Research Scientist

Cengiz Öztireli

Rendering: Computing Integrals

Numerical Integration

Approximate integrals with weighted sum of samples I := 1 |D| Z

D

f(x)dx ˆ I :=

n

X

I=1

wif(xi) biasP[ˆ I] = I − EP[ˆ I] varP[ˆ I] = EP[ˆ I2] − (EP[ˆ I])2

Numerical Integration

Approximate integrals with weighted sum of samples

Random Arrangement Density

Stochastic Point Processes

Formal characterization of point patterns

Stochastic Point Processes

Formal characterization of point patterns

Point Process

Stochastic Point Processes

Examples of point processes

Natural Process Manuel Process

General Point Processes

Infinite point processes

Observation window

General Point Processes

Assign a random variable to each set

B

NpBq “ 3

B

NpBq “ 5

B

NpBq “ 2

General Point Processes

Joint probabilities define the point process

B1 B1 B1 B2 B2 B2

pNpB1q,NpB2q

Point Process Statistics

First order product density

x

Expected number of points around x Measures local density

%(1)(x) = (x)

Point Process Statistics

First order product density

) = (x)

Point Process Statistics

First order product density

) = (x)

Constant

Expected number of points around x & y Measures the joint probability p(x, y)

x y %(2)(x, y) = %(x, y)

Point Process Statistics

Second order product density

Expected number of points around x, y, z

x y

Point Process Statistics

Higher order product density?

z

Not necessary: second order dogma

Point Process Statistics

Summary: 1st & 2nd order correlations sufficient

x x y

Stationary Point Processes

Stationary (translation invariant) Isotropic (translation & rotation invariant)

Stationary Point Processes

Stationary (translation invariant)

λ(x) = λ %(x, y) = %(x − y) = λ2g(x − y)

Pair Correlation Function (PCF) DoF reduced from 2d to d

Stationary Point Processes

Isotropic point process (translation & rotation invariant)

λ(x) = λ g(x − y) = g(||x − y||)

PCF

r

gprq

Estimating Correlations

Second order stationary - pair correlation function (PCF) ˆ g(r) = 1 Kλ2 X

Pk

X

xi,xj2Pk,i6=j

δ(r − (xi − xj))

xi

xj

P1

P2

P3

P4

ˆ g(r) = 1 Kλ2aID(r) X

Pk

X

xi,xj2Pk,i6=j

δ(r − (xi − xj))

Estimating Correlations

Second order stationary - pair correlation function (PCF)

xi

xj

P1

P2

P3

P4

Estimating Correlations

Second order stationary - pair correlation function (PCF)

Pair Correlation Function Point Distribution

ˆ g(r) = 1 λ2rd1|Sd| X

i6=j

k(r kxi xjk) Estimating Correlations

Second order isotropic - pair correlation function (PCF)

Kernel e.g. Gaussian Volume of the unit hypercube in d dimensions

Pair Correlation Function

1

ˆ g(r) = ˆ g(r) =

Pair Correlation Function

1

ˆ g(r) = ˆ g(r) =

Spectral Statistics

Power spectrum Fourier transform

• f PCF

Spectral Statistics sm = X e−i2πmT xj

s(x) = X δ(x − xj)

pm = EP[s∗

msm] = λgm + 1

Spectral Statistics

PCF Power spectrum Points Points

pm = EP[s∗

msm] = λgm + 1

Spectral Statistics

Power spectrum Radial average Radial anisotropy

pm = EP[s∗

msm] = λgm + 1

Statistics for Stationary Processes

Summary Stationary: Spatial (PCF) & spectral (power spectrum) Isotropic: radial averages

PCF Power spectrum

EP 2 4X

I6=j

f(xi, xj) 3 5 = Z

Rd⇥Rd f(x, y)%(x, y)dxdy

(xi

Error in Numerical Integration

Campbell’s Theorem

EP hX f(xi) i = Z

Rd f(x)λ(x)dx

Error in Numerical Integration

Campbell’s theorem for the error of the integral estimator

ˆ I :“

n

ÿ

“

wifpxiq biaspˆ Iq “ I ´ Eˆ I varpˆ Iq “ Eˆ I2 ´ pEˆ Iq2 Eˆ I “ wi “ wpxiq Eˆ I “ E ÿ wpxiqfpxiq “

Campbell’s theorem

q “ ª

V

wpxqfpxqλpxqdx wpxq “ 1{λpxq Ñ biaspˆ Iq “ 0

Eˆ I2

Error in Numerical Integration

Campbell’s theorem for the error of the integral estimator

ˆ I :“

n

ÿ

“

wifpxiq biaspˆ Iq “ I ´ Eˆ I varpˆ Iq “ Eˆ I2 ´ pEˆ Iq2 “ ª

V ˆV

wpxqfpxqwpyqfpyq%px, yqdxdy ` ª

V

w2pxqf 2pxqpxqdx Eˆ2 “ E ÿ

i‰j

wifiwjfj ` E ÿ pwifiq2

var(ˆ I) = 1 λ Z f 2(x)dx + Z af(r)g(r)dh − ✓Z f(x)dx ◆2

Error in Numerical Integration

Stationary point processes

λ(x) = λ %(x, y) = % λ2g(x − y)

Density Arrangement

biaspˆ Iq “ 0 pw “ 1{λq g “ 1 g ° 1 g † 1

var(ˆ I) = 1 λ Z f 2(x)dx + Z af(r)g(r)dh − ✓Z f(x)dx ◆2

Error in Numerical Integration

Stationary point processes

Z f(x)d

Z af(r)g

r)g(r)d

Z af(r)g(r)d

Error in Numerical Integration

Importance Sampling – invertible warp

λ, g(r)

(x), %(x, y)

Error in Numerical Integration

Importance Sampling – general unbiased

var(ˆ I) = Z f 2(x) (x) dx + Z f(x)f(y) %(x, y) (x)(y)dxdy − ✓Z f(x)dx ◆2

Importance Sampling – random add/remove for intensity

var(ˆ I) = Z f 2(x) λ(x) dx + Z f(x)f(y)g(x − y)dxdy − ✓Z f(x)dx ◆2

Error in Numerical Integration

Spectral counterparts

var(ˆ I) = I2var(s0) + X

m6=0

f ⇤

mfmE[s⇤ msm] +

X

l6=m

f ⇤

mflE[sms⇤ l ]

s(x) = X w(xi)δ(x − xi)

sm = X w(xi)e−2πmT xi

Phase Non-stationary Power spectra Stationary

Error in Numerical Integration

Spectral counterparts – stationary point processes

var(ˆ I) =

s(x) = X w(xi)δ(x − xi)

sm = X w(xi)e−2πmT xi

Power spectra Stationary

) + X

m6=0

f ⇤

mfmE[s⇤ msm] +

Error in Numerical Integration

Spectral counterparts – stationary point processes

s(x) = X w(xi)δ(x − xi)

sm = X w(xi)e−2πmT xi

var(ˆ I) =

X

m6=0

(f ⇤

mfm)pm

pm = E[s∗

msm] = λgm + 1

Error in Numerical Integration

Stationary point processes – spatial vs. spectral

var(ˆ I) =

1 λ Z f 2(x)dx + Z af(r)g(r)dh − ✓Z f(x)dx ◆2

X

m6=0

(f ⇤

mfm)pm

2nd order pair-wise correlations 1st order correlations by warp/ algorithm Study in spatial or spectral

Analysis of Sample Correlations for Monte Carlo Rendering

SAMPLING MEASURES & ERROR FORMULATIONS

Research Scientist

Cengiz Öztireli