Light, Reflectance,

and Global Illumination

TOPICS:
Survey of Representations for Light and Visibility
Color, Perception, and Light
Reflectance
Cost of Ray Tracing
Global Illumination Overview
Radiosity
Yus Inverse Global Illumination paper

15-869, Image-Based Modeling and Rendering

Paul Heckbert, 18 Oct. 1999

Representations for Light and Visibility

fixed viewpoint or
planar scene

2-D

diffuse surface

2.5-D/ textured range data (z,r,g,b)

3-D
or pre-lit polygons

specular surface in
clear air, unoccluded

4-D

specular surface in fog, 5-D

or occluded.
plenoptic function I(x,y,z,,)

image (r,g,b)

light field
5-D light field
or surface/volume model with
general reflectance properties

italics: not image-based technique

Physics of Light and Color

Light comes from electromagnetic radiation.
The amplitude of radiation is a function of wavelength .
Any quantity related to wavelength is called spectral.
Frequency = 2 /
EM spectrum:
radio, infrared, visible, ultraviolet, x-ray, gamma-ray
long wavelength
short wavelength
low frequency
high frequency
Dont confuse EM ``spectrum with ``spectrum of signal in signal/image processing;
theyre usually conceptually distinct.
Likewise, dont confuse spectral wavelength & frequency of EM spectrum with spatial
wavelength & frequency in image processing.

Perception of Light and Color

Light is electromagnetic radiation visible to the human eye.
Humans have evolved to sense the dominant wavelengths emitted by
the sun; space aliens have presumably evolved differently.
In low light, we perceive primarily with rods in the retina, which
have broad spectral sensitivity (at night, you cant see color).
In bright light, we perceive primarily with the three sets of cones in
the retina, which have narrower spectral sensitivity roughly
corresponding to red, yellow, and blue wavelengths.
Were most sensitive to greens & yellows; not very sens. to blue.
The eye has a huge dynamic range: 10 orders of magnitude.
In the brain, neurons combine cone signals into spot, edge, and line
receptive fields (point spread functions). And each comes at a
range of scales, orientations, and color channels.

What is Color?
Color is human perception of light.
Our perception is imperfect; we dont see spectral power P(),
instead we see three scalars:
at long wavelength ( red):
L = P()sL()d
at middle wavelength ( yellow):

M = P()sM()d

at short wavelength ( blue):

S = P()sS()d

where sL(), sM(), and sS() are the sensitivity curves of our three
sets of cones.
Color is three dimensional because any light we see is
indistinguishable to our eyes from some mixture of three spectral
(monochromatic, single-wavelength) primaries.

Color Spaces
Color can be described in various color spaces:
spectrum - allows non-visible radiation to be described, but usually
impractical/unnecessary.
RGB - CRT-oriented color space
good for computer storage.

HSV - a more intuitive color space; good for user interfaces

H=hue -- the color wheel, the spectral colors
S=saturation (purity) -- how gray?
V=value (related to brightness, luminance) -- how bright?
a non-linear transform of RGB, since H is cyclic.

CIE XYZ - used by color scientists, a linear transform of RGB.

other color spaces, less commonly used...
For extremely realistic image synthesis, use of four or more samples of the spectrum may be
necessary, but for most purposes, the three samples used by RGB color space is just fine.

Light is a Function of Many Variables

Light is a function of
position
x,y,z
direction
wavelength
time t
polarization
phase
In computer graphics, we typically ignore the last three by assuming
static scenes, unpolarized, incoherent light, and assume that the speed
of light is infinite.
But light is still a complicated function of many variables.
How do we measure light, what are the units?

Units of Light
quantity

dimension

units

solid angle

solid angle

[steradian]

power energy/time [watt]=[joule/sec]

radiance L power/(area*solid angle)

a.k.a. intensity I

[watt/(m2*steradian)]

In vacuum or as an approximation for air, radiance is constant along

a ray.
A picture is an array of incoming radiance values at imaginary
projection plane; because of radiance-constancy, these are equal to
outgoing radiances at intersection of ray with first surface hit
In general, light is absorbed and scattered along a ray.

General Reflection & Transmission

At a surface, reflectance is the fraction of incident (incoming) light that
is reflected, transmittance is the fraction that is transmitted into the
material. Opaque materials have zero transmittance.
In some books, reflectance is denoted , and transmittance . Other
books use kdr and ksr for diffuse and specular reflectance, and kdt and
kst for transmittance.
A general reflectance function has the form of a bidirectional
reflectance distribution function (BRDF): (i,i,o,o), where the
direction of incoming light is (i,i) and the direction of outgoing light
is (o,o), is the polar angle measured from perpendicular, and is
the azimuth.
There is a similar function for bidirectional transmittance, (i,i,o,o).
Light is absorbed and scattered by some media (e.g. fog).
Phongs Illumination model is an approximation to general reflectance.

Phong Illumination Model

A point light source with radiance Il, illuminating an opaque surface,
reflects light of the following radiance:
If surface is perfectly diffuse (Lambertian). It is independent of viewing direction!
I = Il kdr max{N.L,0}/r2
where kdr = coefficient of diffuse reflection [1/steradian]
N = unit normal vector
L = unit direction vector to light, r is distance to light
If surface is perfectly specular. It is not independent of viewing direction.
I = Il ksr max{N.H,0}e/r2
where ksr = coefficient of specular reflection [1/steradian]
H = unit halfway vector = (V+L)/||V+L||
V = unit direction vector to viewer
e = exponent, controlling apparent roughness: small=rough, big=smooth
There are more realistic reflection models than Phongs.
Dont confuse Phong Illumination with Phong Shading.

Cost of Basic Rendering Algorithms

s = #surfaces (e.g. polygons)

ts = time per surface (transforming, ...)

p = #pixels

tp = time per pixel (writing, incrementing, ...)

= #lights

t = time to light surface point w.r.t. one light

a = screen areas of surfs

ti = time for one ray/surface intersection test

Painters or Z-buffer algorithm, with flat shading

(assuming no sorting in painters algorithm)
worst case cost = s(ts + t) + atp atp if polygons big

Painters or Z-buffer algorithm, with per pixel shading (e.g. Phong)

worst case cost = sts + a(t+ tp)

at if polygons big

Ray casting with no shadows, no spatial data structures

worst case cost = p(sti + t + tp)

psti if many surfaces

Ray tracing to max depth d with shadows, refl&tran, no spat. DS, no supersampling
2d1 intersections/pixel, for each of which there are shadow rays
worst case cost = p(2d1)[( +1)sti + t] 2dpsti if many surfaces
Note: time constants vary, e.g. tp is larger for z-buffer than for painters.

Interreflection
We typically simulate just direct illumination: light traveling on a
straight, unoccluded line from light source to surface, reflected
there, then traveling in a straight, unoccluded line into eye.
Light travels by a variety of paths:
light source eye (0 bounces: looking at light source)
light source surface1 eye
(1 bounce: direct illumination)
light source surf1 surf2 eye
(2 bounces)
light source surf1 surf2 surf3 eye
(3 bounces)
...
Illumination via a path of 2 or more bounces is called indirect
illumination or interreflection. It also happens with transmission.

Global Illumination
Observation: light comes from other surfaces, not just designated light
sources.
Goal: simulate interreflection of light in 3-D scenes.
Difficulty: you can no longer shade surfaces one at a time, since theyre
now interrelated!
Two general classes of algorithms:
1. ray tracing methods: simulate motion of photons one by one, tracing
photon paths either backwards (eye ray tracing) or forwards (light ray
tracing) -- good for specular scenes
2. radiosity methods: set up a system of linear equations whose solution is
the light distribution -- good for diffuse scenes

The Unit of Radiosity

Radiance (a.k.a. intensity) is power from/to an area in a given
direction.
units: power / (area solid angle)

Radiosity is outgoing power per unit area due to emission or reflection

over a hemisphere of directions.
units: power / area
radiosity = radiance integral of [cos(polar angle) d(solid angle)] over a
hemisphere = radiance
So radiosity and radiance are linearly interrelated.
Thus, radiosity is both a unit of light and an algorithm.

Radiant emitted flux density is the unit for light emission.

units: power / area

Radiosity as an Integral Equation

radiosity(x) emitted(x) reflectance(x) formfactor(x,t) radiosity(t) dt

where x and t are
surface points

other
surface
pointst

This is called an integral equation because the unknown function

radiosity(x) appears inside an integral.
Can be solved by radiosity methods or randomized Monte Carlo
techniques also, by simulating millions of photon paths.
Radiosity methods are a discrete way to think about and simulate
global illumination.

Classical Radiosity Method

Definitions:
surfaces are divided into elements
radiosity = integral of emitted radiance plus reflected radiance over a
hemisphere. units: [power/area]

Assumptions:
no participating media (no fog) shade surfaces only, not vols.
opaque surfaces (no transmission)
reflection and emission are diffuse radiance is direction-indep.,
radiance is a function of 2-D surface parameters and
reflection and emission are independent of within each of several
wavelength bands; typically use 3 bands: R,G,B solve 3 linear systems
of equations
radiosity is constant across each element one RGB radiosity per element

Typically (but not exclusively):

surfaces are polygons, elements are quadrilaterals or triangles

Deriving Radiosity Equations, 1

outgoing
power

ofelemi

power power
emitted reflected
byelemi byelemi

outgoing
power

ofelemi

power
fractionofpower outgoing

reflectance
leavingelemjthat power
emitted

ofelemi

byelemi
elem j arrivesatelemi ofelemj

Let
Ai = areaofelementi(computable)
ei = radiantemittedfluxdensityofelementi(given)
i = reflectanceofelementi(given)
bi = radiosityofelementi(unknown)
Fij = formfactorfromitoj = fractionofpowerleavingithatarrivesatj
(computable)
Sotheequationabovecanberewritten:
n

Ai bi Ai ei i

F
j 1

ji A j b j

Form Factors
Define the form factor Fij to be the fraction of light leaving element i
that arrives at element j
cosi cos j
1
Fij
vij dA j dAi
2
r
Ai
Ai Aj
Where

vij is a boolean visibility function: 0 if point on i is occluded with respect to

point on j, 1 if unoccluded.
This is a double area integral. Difficult! We end up approximating it.
dAi and dAj are infinitesimal areas on elements i and j, respectively
i and j are polar angles: the angles between ray and normals on elements i and
j, respectively
Projected area of dAi from j is cos i dAi, hence the cosines
r is distance from point on i to point on j

Reciprocity law: AiFij = AjFji.

Deriving Radiosity Equations, 2

n

Earlier, wehad : Ai bi Ai ei i
n

DividingbyAi : bi ei i

Fji A j b j
j1

Aj

Fji Ai b j
j1

Bythereciprocitylaw,Fji ( A j / Ai ) Fij ,sobi ei i

Fij b j forallelemsi

j1

Or, inmatrix/vectornotation:
b = e + Kb
wherebisthevectorofunknownradiosities,eisthevectorofknownemissions,
andKisasquarematrixofreflectancetimesformfactor: K ij i Fij
SubtractingKbfrombothsides,wegetb Kb = e, or
(I K)b = e
whereIisanidentitymatrix.
Thisisalinearsystemofnequationsinnunknowns(thebi ).
Therearethreesuchsystemsofequations,onefortheredchannel,onefor
green, andoneforblue.Thevariables i , ei , andbi areRGBvectors.

Computing Visibility for Form Factors

Computing visibility in the form factor integral is like solving a
hidden surface problem from the point of view of each surface in
the scene.
Two methods:
ray tracing: easy to implement, but can be slow without spatial subdivision
methods (grids, octrees, hierarchical bounding volumes) to speed up raysurface intersection testing
hemicube: exploit speed of z-buffer algorithm, compute visibility between one
element and all other elements. Good when you have z-buffer hardware, but
some tricky issues regarding hemicube resolution
You end up approximating the double area integral with a double summation, just
like numerical methods for approximating integrals.
When two elements are known to be inter-visible (no occluders), you can use
analytic form factor formulas and skip all this.

Radiosity Systems Issues

All algorithms require the following operations:
1. Input scene (geometry, emissions, reflectances).
2. Choose mesh (important!), subdividing polygons into elements.
3. Compute form factors using ray tracing or hemicube for visibility (expensive).
4. Solve system of equations (indirectly, if progressive radiosity).
5. Display picture.
If mesh is chosen too coarse, approximation is poor, you get blocky shadows.
If mesh is chosen too fine, algorithm is slow. A good mesh is critical!
In radiosity simulations, because scene is assumed diffuse, surfaces radiance will be
view-independent.
Changes in viewpoint require only visibility computations, not shading. Do with zbuffer hardware for speed. This is commonly used for architectural
walkthroughs and virtual reality.
Changes in scene geometry or reflectance require a new radiosity simulation.

Summary of Radiosity Algorithms

Radiosity algorithms allow indirect lighting to be simulated.

Classical radiosity algorithms:

Generality: limited to diffuse, polygonal scenes.
Realism: acceptable for simple scenes; blocky shadows on complex scenes. Trial
and error is used to find the right mesh.
Speed: good for simple scenes. If all form factors are computed, O(n2), but if
progressive radiosity or newer hierarchical radiosity algorithms are used,
sometimes O(n).

Generalizations:
curved surfaces: easy - radiosity samples are like a surface texture.
non-diffuse (specular or general) reflectance: much harder; radiosity is now a
function of not just 2-D position on surface, but 2-D position and 2-D direction.
Lots of memory required, but it can be done.

Inverse Global Illumination:

Recovering Reflectance Models of
Real Scenes from Photographs
Yizhou Yu, Paul Debevec, Jitendra Malik, and Tim Hawkins
SIGGRAPH 99

15-869, Image-Based Modeling and Rendering

Paul Heckbert, 20 Oct. 1999

Yu: Motivation
Most IBMR methods permit novel viewpoints but not novel lighting;
they assume surfaces are diffuse.
Want to permit changes in lighting.
Would like to to recover (non-diffuse) reflectance maps:
specularity and roughness at each surface point

Yu: Simplifying Assumptions

static scene of opaque surfaces

known shape
known light source positions
calibrated cameras, known camera positions
high dynamic range photographs
specular reflection parameters (specular reflection coefficient and
roughness) constant over large surface regions
each surface point captured in at least one image
each light source captured in at least one image
image of a highlight in each specular surface region in at least one
photograph

Yu Algorithm Overview
1. Solve for shape (use FAADE or similar system)
2. Solve for coarse diffuse reflectances
coarsely mesh the surfaces into triangles
inverse radiosity problem: known form factors and radiosities (from image
radiances), solve for diffuse reflectances (linear system)

3. Solve for coarse specular reflectance

find highlights for each surface from known surface geometry & lights
check that highlight unoccluded

sample image at points around the highlight center

repeat until convergence
solve for diffuse and specular reflectances (linear system) for each region
solve for roughness of each region (nonlinear)
update estimates of interreflection between surfaces

4. Solve for detailed diffuse reflectance (albedo maps)

look at all images of a given surface point, subtract specular component
throw out outlier (highlight-tainted values), average the rest

Yu: Acquisition Details

spherical, frosted light bulbs

180 volt DC power to avoid 60Hz light flicker
black strips of tape on walls for digitization
shot 150 images with digital camera, merged to create 40 high
dynamic range images (represent with floating point pixel values)
block light source to camera in some of the images
correct for radial distortion and vignetting