Global Illumination

CSE167: Computer GraphicsInstructor: Steve Rotenberg

UCSD, Fall 2006

Classic Ray Tracing The ‘classic’ ray tracing algorithm shoots one primary ray per pixel If the ray hits a colored surface, then a shadow ray is shot towards each light

source to test for shadows, and determine if the light can contribute to the illumination of the surface

If the ray hits a shiny reflective surface, a secondary ray is spawned in the reflection direction and recursively traced through the scene

If a ray hits a transparent surface, then a reflection and a transmission (refraction) ray are spawned and recursively traced through the scene

To prevent infinite loops, the recursion depth is usually capped to some reasonable number of bounces (less than 10 usually works)

In this way, we may end up with an average of fewer than 20 or so rays per pixel in scenes with only a few lights and a few reflective or refractive surfaces

Scenes with many lights and many inter-reflecting surfaces will require more rays Images rendered with the classic ray tracing algorithm can contain shadows,

exact inter-reflections and refractions, and multiple lights, but may tend to have a rather ‘sharp’ appearance, due to the limitation to perfectly polished surfaces and point light sources

Classic Ray Tracing

etc.

Distribution Ray Tracing Distribution ray tracing extends the classic ray tracing algorithm by shooting several

rays in situations where the classic algorithm shoots only one (or two) For example, if we shoot several primary rays for a single pixel, we can achieve

image antialiasing We can model area light sources, and achieve soft edge shadows by shooting

several shadow rays distributed across the light surface We can model blurry reflections and refractions by spawning several rays distributed

around the reflection/refraction direction We can also model camera focus blur by distributing our rays across a virtual camera

aperture As if that weren’t enough, we can also render motion blur by distributing our primary

rays in time Distribution ray tracing is a powerful extension to classic ray tracing that clearly

showed that the central concept of ray tracing was a useful paradigm for high quality rendering

However, it is, of course, much more expensive, as the average number of rays per pixel can jump to hundreds, or even thousands…

Distribution Ray Tracing

etc.

etc.

Ray Tracing The classic and distribution ray tracing algorithms are clearly

important steps in the direction of photoreal rendering However, they are not truly physically correct as they still are

leaving out some components of the illumination In particular, they don’t fully sample the hemisphere of possible

directions for incoming light reflected off of other surfaces This leaves out important lighting features such as color bleeding

also known as diffuse inter-reflection (for example, if we have a white light source and a diffuse green wall next to a diffuse white wall, the white wall will appear greenish near the green wall, due to green light diffusely reflected off of the green wall)

It also leaves out complex specular effects like focused beams of light known as caustics (like the wavy lines of light seen at the bottom of a swimming pool)

Hemispherical Sampling We can modify the distribution ray tracing algorithm to shoot a bunch of

rays scattered about the hemisphere to capture additional incoming light With some careful tuning, we can make this operate in a physically plausible

way However, we would need to shoot a lot of rays to adequately sample the

entire hemisphere, and each of those rays would have to spawn lots of other rays when they hit surfaces

10 rays is definitely not enough to sample a hemisphere, but let’s just assume for now that we will use 10 samples for each hemisphere

If we have 2 lights and we supersample the pixel with 16 samples and allow 5 bounces where each bounce shoots 10 rays, we end up with potentially 16*(2+1)*105 = 4,800,000 rays traced to color a single pixel

This makes this approach pretty impractical The good news is that there are better options…

Path Tracing In 1985, James Kajiya proposed the Monte Carlo path tracing algorithm, also known as

MCPT or simply path tracing The path tracing algorithm fixes many of the exponential ray problems we get with

distribution ray tracing It assumes that as long as we are taking enough samples of the pixel in total, we

shouldn’t have to spawn many rays at each bounce Instead, we can even get away with spawning a single ray for each bounce, where the

ray is randomly scattered somewhere across the hemisphere For example, to render a single pixel, we may start by shooting 16 primary rays to

achieve our pixel antialiasing For each of those samples, we might only spawn off, say 10 new rays, scattered in

random directions From then on, any additional bounces will spawn off only 1 new ray, thus creating a

path. In this example, we would be tracing a total of 16*10 paths per pixel We will still end up shooting more than 160 rays, however, as each path may have

several bounces and will also spawn off shadow rays at each bounce Therefore, if we allow 5 bounces and 2 lights, as in the last example, we will have a

total of (2+1)*(5+1) = 18 rays per path, for a total of 8*160=1280 rays per pixel, which is a lot, but far more reasonable than the previous example

Path Tracing

BRDFs In a previous lecture, we briefly introduced the concept of a BRDF,

or bidirectional reflectance distribution function The BRDF is a function that describes how light is scattered (reflected) off

of a surface The BRDF can model the macroscopic behavior of microscopic surface

features such as roughness, different pigments, fine scale structure, and more

The BRDF can provide everything necessary to determine how much light from an incident beam coming from any direction will scatter off in any other direction

Different BRDFs have been designed to model the complex light scattering patterns from a wide range of materials including brushed metals, human skin, car paint, glass, CDs, and more

BRDFs can also be measured from real world materials using specialized equipment

BRDF Formulation The wavelength dependent BRDF at a point is a 5D function

BRDF = fr(θi,φi,θr,φr,λ)

Often, instead of thinking of it as a 5D scalar function of λ, we can think of it as a 4D function that returns a colorBRDF = fr(θi,φi,θr,φr)

Another option is to express it in more of a vector notation:BRDF = fr(ωi,ωr)

Sometimes, it is also expressed as a function of position:BRDF = fr(x,ωi,ωr)

Physically Plausible BRDFs For a BRDF to be physically plausible, it must not violate two key laws of physics:

Helmholtz reciprocityfr(ωi,ωr) = fr(ωr,ωi)

Helmholtz reciprocity refers to the reversibility of light paths. We should be able to reverse the incident and reflected ray directions and get the same result. It is this important property of light that makes algorithms like ray tracing possible, as they rely on tracing light paths backwards

Conservation of energy

∫Ω fr(ωi,ωr)(ωr·n)dωr < 1, for all ωi

For a BRDF to conserve energy, it must not reflect more light than it receives. A single beam of incident light may be scattered across the entire hemisphere above the surface. The total amount of this reflected light is the (double) integral of the BRDF over the hemisphere of possible reflection directions

BRDF Evaluation The outgoing radiance along a vector ωr due to an

incoming radiance (irradiance) from direction ωi:

dLr(x,ωr)=fr(x,ωi,ωr)Li(x,ωi)(ωi·n)dωi

To compute the total outgoing radiance along

vector ωr, we must integrate over the hemisphere of incoming radiance:

Lr(x,ωr)=∫Ω fr(x,ωi,ωr)Li(x,ωi)(ωi·n)dωi

Rendering EquationLr(x,ωr)=∫Ω fr(x,ωi,ωr)Li(x,ωi)(ωi·n)dωi

This equation is known as the rendering equation, and is the key mathematical equation behind modern photoreal rendering

It describes the light Lr reflected off from some location x in some direction ωr For example, if our primary ray hits some surface, we want to know the light reflected off of

that point back in the direction towards the camera The reflected light is described as an integral over a hemispherical domain Ω, which is really

just shorthand for writing it as a double integral over two angular variables We integrate over the hemisphere of possible incident light directions ωi Given a particular incident light direction ωi and our desired reflection direction ωr, we

evaluate the BRDF fr() at location x The BRDF tells us how much the light coming from direction ωi will be scaled, but we still

need to know how much light is coming from that direction. Unfortunately, this involves computing Li(), which involves solving an integral equation exactly like the one we’re already trying to solve

The rendering equation is unfortunately, an infinitely recursive integral equation, which makes it rather difficult to compute

Monte Carlo Sampling Path tracing is based on a mathematical concept of Monte Carlo sampling Monte Carlo sampling refers to algorithms that make use of randomness to compute

a mathematical result (Monte Carlo famous for its casinos) Technically, we use Monte Carlo sampling to approximate a complex integral that we

can’t solve analytically For example, consider computing the area of a circle. Now, we have a simple

analytical formula for that, but we can apply Monte Carlo sampling to it anyway We consider a square area around our circle and choose a bunch of random points

distributed in the square. If we count the number of points that end up inside the circle, we can approximate the area of the circle as:

(area of square) * (number of points in circle) / (total number of points)

Monte Carlo sampling is a ‘brute force’ computation method for approximating complex integrals that can’t be solved with any other reasonable way. It is often considered as a last resort to solving complex problems, as it can at least try to approximate any integral equation, but it may require lots of samples

Limitations of Path Tracing Path tracing can be used to render any lighting effect,

but might require many paths to adequately resolve complex situations

Some situations are simply too complex for the algorithm and would require too many paths to make it practical

It is particularly bad at capturing specularly bounced light and focused beams of light

Path tracing will converge on the ‘correct’ solution, given enough paths, so it can be used to generate reference images of how a scene should look. This can be useful for evaluating and comparing with other techniques

Bidirectional Path Tracing Bidirectional path tracing (BPT) is an extension to the basic path

tracing algorithm that attempts to handle indirect light paths better For each pixel, several bidirectional paths are examined We start by tracing a ray path from the eye as in path tracing. We

then shoot a photon path out from one of the light sources Lets say each path has 5 rays in it and 5 intersection points We then connect each of the intersection points in the eye path with

each of the intersection points in the light path with a new ray If the ray is unblocked, we add a contribution of the new path that

connects from the light source to the eye The BPT algorithm improves on path tracing’s ability to handle

indirect light (such as a room lit by light fixtures that shine on the ceiling)

Metropolis Sampling Metropolis sampling is another variation of a path tracing type algorithm For some pixel, we start by tracing a path from the eye to some light source We then make a series of random ‘modifications’ to the path and test the amount of

light from that path Based on a statistical algorithm that uses randomness, a decision is made whether or

not to ‘keep’ the new path or discard it Whichever way is chosen, the resulting path is then modified again and the algorithm

is repeated Metropolis sampling is quite a bizarre algorithm and makes use of some complex

properties of statistics It tends to be good at rendering highly bounced light paths such as a room lit by

skylight coming through a window, or the caustic light patterns one sees at the bottom of swimming pools

The Metropolis algorithm is difficult to implement and requires some very heuristic components

It has demonstrated an ability to render features that very few other techniques can handle, although it has not gained wide acceptance and tends to be limited to academic research

Photon Mapping In 1995, Henrik Jensen proposed the photon mapping algorithm Photon mapping starts by shooting many (millions) of photons from the light sources

in the scene, scattered in random directions Each photon may bounce off of several surfaces in a direction that is random, but

biased by the BRDF of the surface For each hit, a statistical (random) decision is made whether or not the photon will

bounce or ‘stick’ in the surface, eventually leading to all photons sticking somewhere The photons are collected into a 3D data structure (like a KD tree) and stored as a

bunch of 3D points with some additional information (color, direction…) Next, the scene is rendered with a ray/path tracing type approach Rays are shot from the camera and may spawn new rays off of sharp reflecting or

refracting surfaces The more diffuse components of the lighting can come from analyzing the photon

map To compute the local lighting due to photons, we collect all of the photons within

some radius of our sample point The photons we collect are used to contribute to the lighting of that point

Photon Mapping

Photon Mapping There are many variations on the photon mapping algorithm and

there are different ways to use the photon map The technique can also be combined with other approaches like

path tracing Photon mapping tends to be particularly good at rendering caustics,

which had previously been very difficult As the photon mapping algorithm stores the photons as simple

points in space, the photon map itself is independent of the scene geometry, making the algorithm very flexible and able to work with any type of geometry

Some of the early pictures of photon mapping showed focused caustics through a glass of sherry, shining onto a surface of procedurally generated sand- a task that would have been impossible with any previous technique

Volumetric Photon Mapping A later extension to the core photon mapping algorithm allowed

interactions between photons and volumetric objects like smoke A photon traveling through a participating media like smoke will

travel in a straight line until it happens to hit a smoke particle, then it will scatter in some random direction, based on the scattering function of the media

This can be implemented by randomly determining the point at which the photon hits a particle (or determining that it doesn’t hit a anything, allowing it to pass through)

As with surfaces, the photons will also be able to ‘stick’ in the volume

Volumetric photon mapping allows accurate rendering of beams of light (even beams of reflected or focused light) to properly scatter through smoke and fog, with the scattered light properly illuminating other objects nearby

Translucency The photon mapping

algorithm (and other ray based rendering algorithms) has also been adapted to trace the paths of photons scattered through translucent surfaces

Local vs. Global Illumination High quality photoreal rendering requires two key things:

1. Accurate modeling of light interactions with materials (local illumination)2. Accurate modeling of light bouncing around within an environment

(global illumination)

The local illumination is modeled with the use of BRDFs. Various BRDF functions have been designed to model the appearance of a variety of materials. BRDFs can also be scanned from real-world objects

The global illumination is computed by tracing light paths in the environment. The various ray-based algorithms take different approaches to evaluating the paths