Global Illumination
Jian Huang, CS 594, Fall 2002
This set of slides reference text book and the
course note of Dutre et. al on SIGGRAPH 2001
Looking Back
Ray-tracing and radiosity both computes
global illumination
Is there a more general methodology?
Its a game of light transport.
Radiance
Radiance (L): for a point in 3D space, L is the light flux per
unit projected area per unit solid angle, measured in W/(srm2)
sr steradian: unit of solid angle
A cone that covers r2 area on the radius-r hemisphere
A total of 2 sr on a hemisphere .
power density/solid angel
The fundamental radiometric quantity
Irradiance and Radiosity
Irradiance (E)
Integration of incoming radiance over all directions, measured in
W/m2
Incident radiant power (Watt) on per unit projected surface area
Radiance distribution is generally discontinuous, irradiance
distribution is generally continuous, due to the integration
shooting, distribute radiance from a surface
gathering, integrating irradiance and accumulate light flux on
surface
Radiosity (B) is
Exitant radiant power (Watt) on per unit projected surface area,
measured in W/m2 as well
Relationships among the
Radiometric Units
Path Notation
A non-mathematical way to categorize the
behavior of global illumination algorithm
Diffuse to diffuse transfer
Specular to diffuse transfer
Diffuse to specular transfer
Specular to specular transfer
Heckberts string notation (1990): as light ray
travels from source (L) to eye (E):
LDDE, LDSE+LDDE, LSSE+LDSE, LSDE, LSSDE
BRDF
Materials interact with light in different ways, and different materials
have different appearances given the same lighting conditions.
The reflectance properties of a surface are described by a reflectance
function, which models the interaction of light reflecting at a surface.
The bi-directional reflectance distribution function (BRDF) is the most
general expression of reflectance of a material
The BRDF is defined as the ratio between differential radiance
reflected in an exitant direction, and incident irradiance through a
differential solid angle
BRDF
The geometry of BRDF
BRDF properties
Positive, and variable in regard to wave-length
Reciprocity: the value of the BRDF will remain
unchanged if the incident and exitant directions
are interchanged.
Generally, the BRDF is anisotropic.
BRDF behaves as a linear function with respect to
all incident directions.
BRDF Examples
Diffuse surface (Lambertian)
Perfect specular surface
BRDF is non-zero in only one exitant direction
Glossy surfaces (non ideally specular)
Difficult to model analytically
Transparent surfaces
Need to model the full sphere (hemi-sphere is not enough)
BRDF is not usually enough, need BSSRDF (bi-directional subsurface scattering reflectance distribution function)
The transparent side can be diffuse, specular or glossy
Reflectance
3 forms
The Rendering Equation
Proposed by Jim Kajiya in his SIGGRAPH1986
paper
Light transport equation in a general form
Describes not only diffuse surfaces, but also ones with
complex reflective properties
Goal of computer graphics: solution of the rendering
equation!
Looks simple and natural, but really is too complex to
be solved exactly; various techniques to nd
approximate solutions are used
The Rendering Equation
I(x,x) = intensity passing from x to x
g(x,x) = geometry term (1, or 1/r2, if x visible
from x, 0 otherwise)
(x,x) = intensity emitted from x in the direction
of x
(x,x,x) = scattering term for x (fraction of
intensity arriving at x from the direction of x
scattered in the direction of x)
S = union of all surfaces
Linear Operator
Define a linear operator, M.
The rendering equation:
How to solve it?
Neumann Series Solution
Start with an initial guess I0
Compute a better solution
Computer an even better solution
Then,
In practice one needs to truncate it somewhere
Examples
No shading/illumination, just draw surfaces
as emitting themselves:
Direct illumination, no shadows:
Direct illumination with shadows:
Implications
How successful is a global illumination algorithm?
The first term is simple, just visibility
How an algorithm handles the remaining terms and the
recursion?
How does it handle the combinations of diffuse and
specular reflectivity
The rendering equation is a view-independent
statement of the problem
How are the radiosity algorithm and the raytracing algorithm?
Monte Carlo Techniques in
Global Illumination
Monte Carlo is a general class of estimation
method based on statistical sampling
The most famous example: to estimate
Monte Carlo techniques are commonly used
to solve integrals with no analytical or
numerical solution
The rendering equation has one such integral
Basic Monte Carlo Integration
Suppose we want to numerically integrate a function over an integration
domain D (of dimension d), i.e., we want to compute the value of the
integral I:
Common deterministic approach: construct a number of sample points,
and use the function values at those points to compute an estimate of I.
Monte Carlo integration basically uses the same approach, but uses a
stochastic process to generate the sample points. And would like to
generate N sample points distributed uniformly over D.
Basic Monte Carlo Integration
The mean of the evaluated function values at each randomly generated
sample point multiplied by the area of the integration domain, provides
an unbiased estimator for I:
Monte Carlo methods provides an un-biased estimator
The variance reduces as N increases
Usually, given the same N, deterministic approach produces less error
than Monte Carlo methods
When to Use Monte Carlo?
High dimension integration the sample points
needed in deterministic approach exponential
increase
Complex integrand: practically cant tell the error
bound for deterministic approaches
Monte Carlo is always un-biased, and for
rendering purpose, it converts errors into noise!!
Two Types of Monte Carlo
Monte Carlo integration methods can roughly be subdivided in
two categories:
those that have no information about the function to be integrated: blind
Monte Carlo
those that do have some kind of information available about the function:
informed Monte Carlo
Intuitively, one expects that informed Monte Carlo methods to
produce more accurate results as opposed to blind Monte Carlo
methods.
The basic Monte Carlo integration is a blind Monte Carlo method
Importance Sampling
An informed Monte Carlo
Importance sampling uses a non-uniform probability
function, pdf(x), for generating samples.
By choosing the probability function pdf(x) wisely on the basis of
some knowledge of the function to be integrated, we can often
reduce the variance
Can prove: if can get the pdf(x) to match the exact shape of the
function to be integrated, f(x), the variance of the integration
estimation is 0.
Practically, can use a sample table to generate a good pdf.
Intuitively, want to send more rays into the more detailed
areas in space
Stratified Sampling
Importance sampling (probability) using a limited number
of samples, which is the case for graphics rendering, does
not have a guarantee.
Stratified sampling address this further: the basic idea of
stratified sampling is to split up the integration domain in
m disjunct subdomains (also called strata), and evaluate
the integral in each of the subdomains separately with one
or more samples.
More precisely:
More On Ray-Tracing
Already discussed recursive ray-tracing!
Improvements to ray-tracing!
Area sampling variations to address aliasing
Cone tracing (only talk about this)
Beam tracing
Pencil tracing
Distributed ray-tracing!
Cone Tracing (1984)
Generalize linear rays into cones
One cone is fired from eye into each pixel
Have a wide angle to encompass the pixel
The cone is intersected with objects in its path
Reflection and refraction are modeled as spherical mirrors
and lenses
Use the curvature of the object intersecting that cone
Broaden the reflected and refracted cones to simulate further
scattering
Shadow: proportion of the shadow cone that remains unblocked
Distributed Ray-Tracing
Another way to address aliasing
By Cook, Porter, and Carpenter in 1984.
A stochastic approach to supersampling that trades
objectionable aliasing artifacts for the less
offensive artifacts of noise
Distributed: rays are stochastically distributed to
sample the quantities
This method was covered during our recursive ray
tracing lecture as extension to correct aliasing
Sampling Other Dimensions
Other than stochastic spatial sampling for anti-aliasing, can
sample in other dimensions
Motion blur (distribute rays in time)
Depth of field (distribute rays over the area of the camera lens)
Rough surfaces: blurred specular reflections and translucent
refraction (distribute rays according to specular reflection and
transmission functions)
Soft shadow: distribute shadow feeler rays over the solid angle
span by the area light source
In all cases, use stochastic sampling to perturb rays
Path Tracing
Devised by Kajiya in 1986
An efficient variation of distributed ray tracing
At each intersection, either fire a refraction ray or a reflection
ray (not both!!)
This decision is guided by the desired distribution of the
different kinds of rays for each pixel
Dont have a binary ray-tree for each initial ray
It is really a Monte Carlo method ! Need to use a quite
large N. (Kajiya used 40 rays per pixel)
Unlike ray-tracing, path tracing traces diffuse rays as well
as specular rays
Path
Global illumination: generate all paths of
light transport that interact among different
surfaces and compute an integration to
solve the rendering equation
Direct paths: length = 1, direct illumination
Indirect paths: length > 1
Indirect paths - surface sampling
Simple generator (path length = 2):
select point on light source
select random point on surfaces
per path:
2 visibility checks
Indirect paths - source shooting
shoot ray from light source, find hit location
connect hit point to receiver
per path:
1 ray intersection
1 visibility check
Indirect paths - receiver gathering
shoot ray from receiver point, find hit location
connect hit point to random point on light source
per path:
1 ray intersection
1 visibility check
Indirect paths
Same principles apply to paths of length > 2
generate multiple surface points
generate multiple bounces from light sources and
connect to receiver
generate multiple bounces from receiver and connect to
light sources
Estimator and noise characteristics change with
path generator
Complex path generators
Bi-directional ray tracing
shoot a path from light source
shoot a path from receiver
connect end points
Combine all paths and weigh
them
Bidirectional ray tracing
Parameters
eye path length = 0: shooting from source
light path length = 0: gathering at receiver
When useful?
Light sources difficult to reach
Specific brdf evaluations (e.g., caustics)
Classic ray tracing?
Classic ray tracing:
shoot shadow-rays (direct illumination)
shoot perfect specular rays only for indirect
ignores many paths
does not solve the rendering equation
