Advanced Computer Graphics

Stochastic Raytracing 1
Matthias Teschner
From Radiosity to Raytracing
− Radiosity equation governs light transport
for diffuse surfaces.  How to describe
light transport for general surfaces?
− How to solve for the light transport?
− How to compute the relevant part of
the light transport towards a sensor?

University of Freiburg – Computer Science Department – 2

◼
Stochastic Raytracing
− Light transport towards the sensor requires to solve

− Monte Carlo integration approximates

the reflectance integral
− E.g.,
− Trace rays into the scene
− Compute radiance along this ray
− Associate an area / solid angle with each ray
− Accumulate all contributions
University of Freiburg – Computer Science Department – 3
◼
Outline
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables

◼University of Freiburg – Computer Science Department – 4

Governing Equations
− Rendering equation
− Governing equation for general global illumination methods

− Radiosity equation
− Governing equation for diffuse global illumination methods

University of Freiburg – Computer Science Department – 5

◼
A Solution Strategy - Radiosity
− Finite Element Method
− Start with a continuous form / function

− Discretization

− Solving for a vector with unknown radiosities

◼University of Freiburg – Computer Science Department – 6

An Alternative Strategy
− Start with the general form of the rendering equation,
e.g. in hemispherical form

− Solving for a function of unknown radiances

− I.e., radiance at all surface positions into all directions

◼University of Freiburg – Computer Science Department – 7

Operator Form of the Rendering Equation
we use operators to shorten the notations

− Operators transform a function into another one

− Scattering operator

− Applied to an incident radiance function , exitant

radiance after one bounce / scattering step is returned

− operates on an entire function, i.e. on all incident

radiances for all positions and direction
See, e.g.: Eric Veach: Robust Monte Carlo Methods for Light
Transport Simulation, Ph.D. thesis, Stanford University, 1997.
◼University of Freiburg – Computer Science Department – 8
Operator Form of the Rendering Equation

− Propagation operator
p’ indicates the raycast operator applied to p

− Applied to an exitant radiance function ,

incident radiance at from direction is returned

− Radiance is preserved / propagated along the line

between and if flux from p to p prime is preserved then flux is preserved in opposite direction also.

− and can be reversed, i.e.

See, e.g.: Eric Veach: Robust Monte Carlo Methods for Light
Transport Simulation, Ph.D. thesis, Stanford University, 1997.
University of Freiburg – Computer Science Department – 9
◼
Operator Form of the Rendering Equation

− Light transport operator

− Composition of scattering and propagation

− Maps an exitant radiance function to the exitant
radiance function after one scattering step
− Remember: maps exitant radiance to incident radiance
propagated along a direction. Then, maps incident
radiance to exitant radiance after scattering

See, e.g.: Eric Veach: Robust Monte Carlo Methods for Light Transport Simulation, Ph.D. dissertation, Stanford University, 1997.

◼University of Freiburg – Computer Science Department – 10

Operator Form of the Rendering Equation

− Can be written as

− Light transport equation

Infinite number of equations with an infinite number of unknown exitant radiances

− relates exitant radiance functions Le adds flux to a scene but T removes flux from a scene

− Represents the light propagation equilibrium

University of Freiburg – Computer Science Department – 11

◼
Light Transport Equation

− Solving for the unknown radiance function

− Neumann series

◼University of Freiburg – Computer Science Department – 12

Light Transport Equation
− Discussion
− Radiance function is a sum of
− Emitted radiance
− Emitted radiance after one scattering
− Emitted radiance after two scatterings
− …

◼University of Freiburg – Computer Science Department – 13

Terms in the Neumann Series
− Example contributions to terms

◼University of Freiburg – Computer Science Department – 14

Forward Raytracing
− Send rays / propagate radiance from all light source
positions into all directions 
− At all intersection points , solve the integral
consider all seen points

for all direction  we will stop the computation when all the sensor or camera elements
are hit by a ray

− Trace rays to propagate

− At all intersection points , solve the integral

for all direction

this is nothing but T Le


◼University of Freiburg – Computer Science Department – 15
 Forward Raytracing
− At a sensor: Accumulate radiance contributions of rays
after n scattering steps, i.e. compute
Light source

accumulating
the radiance
elements

sensor elemnt x receives the

light transport contributions
from a small angle

Sensor
Exemplary rays

◼University of Freiburg – Computer Science Department – 16

Operator Form of the Rendering Equation

− Can be written as

− Light transport equation

Infinite number of equations with an infinite number of unknown exitant radiances

− relates exitant radiance functions

− Represents the light propagation equilibrium

University of Freiburg – Computer Science Department – 17

◼
Light Transport Equation

− Solving for the unknown radiance function

− Neumann series

◼University of Freiburg – Computer Science Department – 18

Light Transport Equation
− Discussion
− Radiance function is a sum of
− Emitted radiance
− Emitted radiance after one scattering
− Emitted radiance after two scatterings
− …

◼University of Freiburg – Computer Science Department – 19

Backward Raytracing
− Consider rays from the sensor into the scene
− Propagate radiance from visible light sources
−  part of visible to the sensor
− At intersection points with the scene, compute
radiance
that is propagated in direction towards the sensor
−  part of visible to the sensor
−…
University of Freiburg – Computer Science Department – 20
◼
Backward Raytracing
− Trace rays from the sensor into the scene

Light source

Sensor
Exemplary rays

◼University of Freiburg – Computer Science Department – 21

Setting at Sensor

Scene
Sensor
Light
source

− How to compute and what is its

the relation to ?
University of Freiburg – Computer Science Department – 22
◼
Setting at First-Level Intersections

this domain splitting is called stratification

light source + scene

− is the part
of visible to the sensor
− Computation of
requires towards
sensor

University of Freiburg – Computer Science Department – 23

◼
Setting at Second-Level Intersections

− is the part of
visible to the sensor
− Computation of
requires towards
p1

University of Freiburg – Computer Science Department – 24

◼
Summary
− Recursive evaluation of

− Each recursion level computes parts of the functions

that are visible to the sensor

University of Freiburg – Computer Science Department – 25

◼
Numerical Integration
− The integral is approximately computed
with a sum of samples
− For each sample i,
− A ray is cast into the scene
− Intersection with the scene is computed
− Radiance along the ray is computed

◼University of Freiburg – Computer Science Department – 26

Numerical Integration
− Typically,
is considered
− For , light source areas are sampled and
rays towards those positions are processed
− For , the respective solid angle is sampled
and rays towards those directions are processed

University of Freiburg – Computer Science Department – 27

◼
Numerical Integration
− Due to the recursive nature, the number of processed
rays grows exponentially with the recursion level
−  Monte Carlo integration
− Efficient for multidimensional integral
− Adaptive sample distribution
− Very flexible in terms of the number of used samples
− Even one sample can be used to approximate an integral
 e.g., Path tracing
− At each recursion level, trace a fixed number of rays to light
sources and one ray into the scene (which generates a ray path)
◼University of Freiburg – Computer Science Department – 28
Advanced Computer Graphics
Stochastic Raytracing 2
Matthias Teschner
Outline
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables

◼University of Freiburg – Computer Science Department – 30

Goal monte carlo goal is to approximate the solution of light transport eqn

− Approximating the solution of the light transport

equation
− Recursive evaluation of

− Each recursion level computes parts of the functions

that are visible to the sensor

◼University of Freiburg – Computer Science Department – 31

Numerical Integration – Fixed Sample Size

− E.g. Riemann sum

−
− More / smaller samples  better accuracy
− dimensional integrals require samples

University of Freiburg – Computer Science Department – 32

◼
Numerical Integration – Adaptive Sample Size

− E.g., Monte Carlo integration

− , adaptive sample size
− More / smaller samples  better accuracy
− dimensional integrals work with arbitrary sample numbers
− Sample size is only approximated  noise

University of Freiburg – Computer Science Department – 33

◼
Stochastic Raytracing - Concept
− Approximately evaluate the reflectance integral
by
− Tracing rays into randomly sampled 2D directions
− Computing the incoming radiances
− Integral is approximated with
we do not know the shape of the angle , we only have the
magnitude of the angle.

− 2 dimensional sample directions

− is an approximation of the solid angle
of sample direction
University of Freiburg – Computer Science Department – 34
◼
Introduction
− Challenges
− Approximate the integral as exact as possible
− Trace as few rays as possible / use as few samples as possible
− Trace relevant rays / use relevant samples
− Rays / samples to light sources are very relevant
(Rays / samples to occluded light sources are irrelevant)
− For diffuse surfaces, rays / samples in normal direction are
more relevant than rays / samples perpendicular to the normal
− For specular surfaces, rays / samples in reflection direction are relevant

◼University of Freiburg – Computer Science Department – 35

Properties
− Benefits
− Processes only evaluations of the integrand
at arbitrary surface points in the domain
− Works for a large variety of integrands,
e.g., it handles discontinuities
− Appropriate for integrals of arbitrary dimensions
− Allows for non-uniform sample patterns /
adaptive sample sizes

University of Freiburg – Computer Science Department – 36

◼
Properties
− Drawbacks
− Using n samples, the scheme converges
to the correct result with O (n½)
− I.e., to half the error, 4n samples are required
− Errors are perceived as noise,
i.e. pixels are arbitrarily too bright or dark
(due to the erroneous approximation of the sample size)
− Evaluation of the integrand at a point and for a direction
is expensive (ray intersection tests)

◼University of Freiburg – Computer Science Department – 37

Continuous Random Variables
− Motivation: random sampling of directions
− Continuous random variables generated by PDF

− In contrast to discrete random variables,

infinite number of possible values
− Canonical uniform random variable
− Sample sets with arbitrary distributions
can be computed from

University of Freiburg – Computer Science Department – 38

◼
Probability Density Function PDF p(x)
also governs the generation of samples

− Motivation: PDF governs the size /

solid angle of a sample / sample direction
− Probability of a random variable taking certain value ranges
−
The probability, that the random
− variable has a certain exact value x =x , is 0.
0 1

− The probability, that the random variable is in the specified domain, is 1.

− Example
− Uniform PDF for
−
◼University of Freiburg – Computer Science Department – 39
Cumulative Distribution Function CDF P(x)

− Motivation: CDFs are required to generate sample sets

for arbitrary PDFs from uniform sample sets
− Probability of a random variable to be less or equal to
−
−
−

University of Freiburg – Computer Science Department – 40

◼
Expected Value of estimator is reflectance integral

− Motivation: expected value of an estimator function

is equal to the reflectance integral
− Expected value of a function is defined as the
weighted average value of the function over a domain
Processes an infinite number of
− with samples x according to a PDF p(x)

− Properties
−
− For independent random variables Xi

◼University of Freiburg – Computer Science Department – 41

Expected Value
− Examples for uniform PDF
−

University of Freiburg – Computer Science Department – 42

◼
Monte Carlo Estimator - Uniform Random Variables

− Motivation: approximation of the reflectance integral

− Goal: computation of
− Uniformly distributed random variables
− Probability density function Constant and integration to one

− Monte Carlo estimator

− Expected value of is equal to the integral
−

◼University of Freiburg – Computer Science Department – 43

Monte Carlo Estimator - Uniform Random Variables

University of Freiburg – Computer Science Department – 44

◼
Monte Carlo Estimator - Uniform Random Variables

− PDF
− Estimator
− Integral
−
− Function value
− Approximate sample size

University of Freiburg – Computer Science Department – 45

◼
Examples - Uniform Random Variables

− Integral
− Estimator Sample size approx. 1/N

− For an increasing number of uniformly distributed

random variables , the
estimator converges to one
−

Uniformly distributed [Suffern]

random samples
◼University of Freiburg – Computer Science Department – 46
Monte Carlo Estimator - Non-uniform Random Variables

− Monte Carlo estimator p (Xi)  0

University of Freiburg – Computer Science Department – 47

◼
Monte Carlo Estimator - Non-uniform Random Variables

− PDF
− Estimator
− Integral
− larger value of p gives more samples in a domain

− Function value
− Approximate sample size

University of Freiburg – Computer Science Department – 48

◼
Approximate Sample Size
constant

− Sample size / distance for uniform PDF:

− Sample size for non-uniform PDF:

linear

University of Freiburg – Computer Science Department – 49

◼
Monte Carlo Estimator - Multiple Dimensions

− E.g.,
− Samples are two-dimensional
− Uniformly distributed random samples

− Probability density function

− Monte Carlo estimator

− Approximate sample size is

◼University of Freiburg – Computer Science Department – 50

Monte Carlo Estimator - Multiple Dimensions

− E.g.,
− Uniformly distributed random samples
− Probability density function
− Monte Carlo estimator
− Approximate sample size

Sample size Sample size

approx. 9/9 approx. 9/14

University of Freiburg – Computer Science Department – 51

◼
Monte Carlo Estimator - Integration over a Hemisphere

− Approximate computation of the irradiance at a point

− Estimator
− Choosing a PDF This flexibility is an important aspect of Monte Carlo integration.

− Should be similar to the shape of the integrand

− As incident radiance is weighted with ,
it is appropriate to generate more samples
close to the top of the hemisphere
−
University of Freiburg – Computer Science Department – 52
◼
Monte Carlo Estimator - Integration over a Hemisphere

− Probability distribution

− Estimator

If  and  are sampled according to PDF p(, )

University of Freiburg – Computer Science Department – 53
◼
Monte Carlo Estimator - Integration over a Hemisphere

− Integral
− PDF
− Estimator

− Function value for direction

− Approximate sample size / solid angle
− For large
− The PDF in terms of the solid angle is
◼University of Freiburg – Computer Science Department – 54
Monte Carlo Integration - Steps
− Choose an appropriate
probability density function
− Generate random samples according to the PDF
− Evaluate the function for all samples
− Accumulate sample values weighted with their
approximate sample size

University of Freiburg – Computer Science Department – 55

◼
Monte Carlo Estimator - Error Reduction

− Importance sampling
− Motivation: contributions of larger
sample values are more important
− PDF should be similar to the shape
of the function
[Suffern]
− Optimal PDF
− E.g., if incident radiance is weighted
with , the PDF should choose
more samples close to the normal
direction
University of Freiburg – Computer Science Department – 56
◼
Monte Carlo Estimator - Error Reduction

− Stratified sampling
− Domain subdivision into strata
− E.g., handling direct and indirect illumination differently
−

light source samples areas

scenes samples solid angle

University of Freiburg – Computer Science Department – 57

◼
Advanced Computer Graphics
Stochastic Raytracing 3
Matthias Teschner
Monte Carlo Estimator - Integration over a Hemisphere

− Approximate computation of the irradiance at a point

− Estimator
− Choosing a PDF This flexibility is an important aspect of Monte Carlo integration.

− Should be similar to the shape of the integrand

− As incident radiance is weighted with ,
it is appropriate to generate more samples
close to the top of the hemisphere
−
University of Freiburg – Computer Science Department – 59
◼
Outline
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables
− Inversion method
− Rejection method
− Transforming between distributions
− 2D sampling
− Examples

University of Freiburg – Computer Science Department – 60

◼
Motivation - Rendering Equation
− Hemispherical form

− Area form

University of Freiburg – Computer Science Department – 61

◼
Motivation - Monte Carlo Integration
− Choose an appropriate
probability density function
− Generate random samples according to the PDF
− Evaluate the function for all samples
− Accumulate function values weighted
with their approximate sample size

University of Freiburg – Computer Science Department – 62

◼
Inversion Method
we have a pdf and the samples are generated based on this pdf

− Mapping of a uniform random

variable to a goal distribution
− Discrete example p1 p2 p3 p4 are relative pbty and their sum is 1

− Four outcomes with probabilities

and
− Computation of the cumulative
distribution function

1 2 3 4
[Pharr, Humphreys]
◼University of Freiburg – Computer Science Department – 63
Inversion Method
− Discrete example cont.
− Take a uniform random variable
− has the desired distribution
− Continuous case
− and are continuous functions 1 2 3 4
− Start with the desired PDF [Pharr, Humphreys]
− Derive
− Compute the inverse
− Obtain a uniformly distributed variable
the pdf . cdf. inversion of cdf are all continuous function

− Compute which adheres to

University of Freiburg – Computer Science Department – 64

◼
Inversion Method - Example 1
− Power distribution used in reflection integral at specular surface

− E.g., for sampling the Blinn microfacet model

− Computation of the PDF
−
− PDF
− CDF
− Inverse of the CDF
− Sample generation
− Generate uniform random samples
− are samples from the distribution
University of Freiburg – Computer Science Department – 65
◼
Inversion Method - Example 2
− Exponential distribution
− E.g., for considering participating media
− Computation of the PDF
−
− PDF
− CDF
− Inverse of the CDF
− Sample generation
− Generate uniform random samples
− are samples from the distribution
◼University of Freiburg – Computer Science Department – 66
Outline
− Context
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables
− Inversion method
− Rejection method
− Transforming between distributions
− 2D sampling
− Examples
University of Freiburg – Computer Science Department – 67
◼
Rejection Method
− Draws samples according to a function
− Dart-throwing approach
− Works with a PDF and a scalar with
− Properties
− is not necessarily a PDF
− PDF, CDF and inverse CDF
do not have to be computed
− Simple to implement
− Useful for debugging purposes

University of Freiburg – Computer Science Department – 68

◼
Rejection Method
− Sample generation
− Generate a uniform random sample
− Generate a sample according to
− Accept if

a b

[Pharr, Humphreys]

◼University of Freiburg – Computer Science Department – 69

Outline
− Context
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables
− Inversion method
− Rejection method
− Transforming between distributions
− 2D sampling
− Examples
University of Freiburg – Computer Science Department – 70
◼
Transforming Between Distributions
− Computation of a resulting PDF, when a function is
applied to samples from an arbitrary distribution
− Random variables are drawn from
− Bijective transformation (one-to-one mapping)
− How does the distribution look like?

◼University of Freiburg – Computer Science Department – 71

Transforming Between Distributions
−

− Example
−
−
−

University of Freiburg – Computer Science Department – 72

◼
Transforming Between Distributions
− Multiple dimensions
− is an n-dimensional random variable
− is a bijective transformation
− Transformation of the PDF

◼University of Freiburg – Computer Science Department – 73

Transforming Between Distributions
− Example (polar coordinates)
− Samples with density
− Corresponding density with and
−

University of Freiburg – Computer Science Department – 74

◼
Transforming Between Distributions
− Example (spherical coordinates)
−
−
−
−
− Example (solid angle)
−
−
−
University of Freiburg – Computer Science Department – 75
◼
Outline
− Context
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables
− Inversion method
− Rejection method
− Transforming between distributions
− 2D sampling
− Examples
University of Freiburg – Computer Science Department – 76
◼
Concept
− Samples from a 2D joint density function
− Procedure
− Compute the marginal density function
− Compute the conditional density function
− Generate a sample according to
− Generate a sample according to
− Marginal density function
− Integral of for a particular over all -values
− Conditional density function
− Density function for given a particular
◼ University of Freiburg – Computer Science Department – 77
Outline
− Context
− Diffuse vs. general global illumination
− Monte Carlo integration
− Sampling of random variables
− Inversion method
− Rejection method
− Transforming between distributions
− 2D sampling
− Examples
University of Freiburg – Computer Science Department – 78
◼
Uniform Sampling of a Triangle
− Sampling an isosceles right triangle of area 0.5
− u, v can be interpreted as Barycentric coordinates
− Can be used to generate samples for arbitrary triangles
−
− Marginal density function (u, 1-u)
− v

− Conditional density u
− Inversion method
−
−
University of Freiburg – Computer Science Department – 79
◼ [Pharr, Humphreys]
Uniform Sampling of a Triangle
− Inversion method cont.
− Inverse functions of the cumulative distribution functions
− u is generated between 0 and 1
− v is generated between 0 and 1-u=(1-)½

− Generating uniformly sampled random values and

− Applying the inverse CDFs to obtain and

University of Freiburg – Computer Science Department – 80

◼
Uniform Sampling of a Hemisphere
− PDF is constant with respect to a solid angle
−
−
− Marginal density function
−
− Conditional density for
−
− Inversion method [Suffern]
−
−
University of Freiburg – Computer Science Department – 81
◼
Uniform Sampling of a Hemisphere
− Inversion method cont.
− Inverse functions of the cumulative distribution functions
−
−
− Generating uniformly sampled random values and
− Applying the inverse CDFs to obtain and
− Conversion to Cartesian space
−
−
−
− is a normalized direction
University of Freiburg – Computer Science Department – 82
◼
Uniform Sampling of a Hemisphere
− Illustration for

Generate less samples

for smaller angles 

◼University of Freiburg – Computer Science Department – 83

Cosine-Weighted Sampling of a Hemisphere

− PDF is proportional to :
−
−
− Marginal density function
−
− Conditional density for
−
− Inversion method
− [Suffern]

−
University of Freiburg – Computer Science Department – 84
◼
Cosine-Weighted Sampling of a Hemisphere

− Inversion method cont.

− Inverse functions of the cumulative distribution functions
−
−
− Generating uniformly sampled random values and
− Applying the inverse CDFs to obtain and
− Conversion to Cartesian space x- y- values uniformly sample a unit
disk, i.e., cosine-weighted samples
− of the hemisphere can also be obtained
− by uniformly sampling a unit sphere
and projecting the samples onto the
− hemisphere
− is a normalized direction
University of Freiburg – Computer Science Department – 85
◼
Cosine-Weighted Sampling of a Hemisphere

− Illustration for

Generate less samples [Suffern]

for smaller and larger
angles  Cosine-weighted
hemisphere
(top view, side view)

◼University of Freiburg – Computer Science Department – 86