MIT6 837F12 Lec18

Download as pdf or txt
Download as pdf or txt
You are on page 1of 84

Global Illumination and Monte Carlo

MIT EECS 6.837 Computer Graphics


Wojciech Matusik
with many slides from Fredo Durand and Jaakko Lehtinen
© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/. 1
Today
• Lots of randomness!

Dunbar & Humphreys 2


Today
• Global Illumination
– Rendering Equation
– Path tracing
• Monte Carlo integration
• Better sampling
– importance
– stratification

© ACM. All rights reserved. This content is excluded from


our Creative Commons license. For more information, see
http://ocw.mit.edu/help/faq-fair-use/.
3
Global Illumination
• So far, we've seen only direct lighting (red here)
• We also want indirect lighting
– Full integral of all directions (multiplied by BRDF)
– In practice, send tons of random rays

4
Direct Illumination

Courtesy of Henrik Wann Jensen. Used with permission. 5


Global Illumination (with Indirect)

Courtesy of Henrik Wann Jensen. Used with permission. 6


Global Illumination
• So far, we only used the BRDF for point lights
– We just summed over all the point light sources
• BRDF also describes how indirect illumination
reflects off surfaces
– Turns summation into integral over hemisphere
– As if every direction had a light source

7
Reflectance Equation, Visually

outgoing light to incident light the BRDF cosine term


direction v from direction
omega

Lin
Lin Lin Sum (integrate)
over every
v
direction on the
hemisphere,
Lin
modulate incident
illumination by
BRDF

8
The Reflectance Equation

• Where does Lin come from?

x
9
The Reflectance Equation

• Where does Lin come from?


– It is the light reflected towards x from the surface point in
direction l ==> must compute similar integral there!
• Recursive!

x
10
The Rendering Equation

• Where does Lin come from?


– It is the light reflected towards x from the surface point in
direction l ==> must compute similar integral there!
• Recursive!
– AND if x happens
to be a light source,
we add its contribution
directly x
11
The Rendering Equation

• The rendering equation describes the appearance of


the scene, including direct and indirect illumination
– An “integral equation”, the unknown solution function L
is both on the LHS and on the RHS inside the integral
• Must either discretize or use Monte Carlo integration
– Originally described by Kajiya and Immel et al. in 1986
– More on 6.839
• Also, see book references towards the end
12
The Rendering Equation
• Analytic solution is usually impossible
• Lots of ways to solve it approximately
• Monte Carlo techniques use random samples for
evaluating the integrals
– We’ll look at some simple method in a bit...
• Finite element methods discretize the solution using
basis functions (again!)
– Radiosity, wavelets, precomputed radiance transfer, etc.

13
Questions?

14
How To Render Global
Illumination?

Lehtinen et al. 2008


© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/. 15
Ray Casting
• Cast a ray from the eye through each pixel

16
Ray Tracing
• Cast a ray from the eye through each pixel
• Trace secondary rays (shadow, reflection, refraction)

17
“Monte-Carlo Ray Tracing”
• Cast a ray from the eye through each pixel
• Cast random rays from the hit point to evaluate
hemispherical integral using random sampling

18
“Monte-Carlo Ray Tracing”
• Cast a ray from the eye through each pixel
• Cast random rays from the visible point
• Recurse

19
“Monte-Carlo Ray Tracing”
• Cast a ray from the eye through each pixel
• Cast random rays from the visible point
• Recurse

20
“Monte-Carlo Ray Tracing”
• Systematically sample light sources at each hit
– Don’t just wait the rays will hit it by chance

21
Results
Henrik Wann Jensen

Courtesy of Henrik Wann Jensen. Used with permission.

22
Monte Carlo Path Tracing
• Trace only one secondary ray per recursion
– Otherwise number of rays explodes!
• But send many primary rays per pixel (antialiasing)

23
Monte Carlo Path Tracing
• Trace only one secondary ray per recursion
– Otherwise number of rays explodes!
• But send many primary rays per pixel (antialiasing)

Again, trace
shadow rays
from each
intersection

24
Monte Carlo Path Tracing
• We shoot one path from the eye at a time
– Connect every surface point on the way to the light by a
shadow ray
– We are randomly sampling the space of all possible light
paths between the source and the camera

25
Path Tracing Results
• 10 paths/pixel
Henrik Wann Jensen

Courtesy of Henrik Wann Jensen. Used with permission. 26


Path Tracing Results: Glossy Scene
• 10 paths/pixel
Note: More noise. This is not a coincidence; the integrand
has higher variance (the BRDFs are “spikier”).
Henrik Wann Jensen

Courtesy of Henrik Wann Jensen. Used with permission.


27
Path Tracing Results: Glossy Scene
• 100 paths/pixel
Henrik Wann Jensen

Courtesy of Henrik Wann Jensen. Used with permission.


28
Importance of Sampling the Light
Without explicit With explicit
light sampling light sampling

1 path
per pixel

4 paths
per pixel


29
Why Use Random Numbers?
• Fixed random sequence
• We see the structure in the error
Henrik Wann Jensen

Courtesy of Henrik Wann Jensen. Used with permission.


30
Demo
• http://madebyevan.com/webgl-path-tracing/

Image removed due to copyright restrictions. Please see the above link for further details.

31
For more demo/experimentation
• http://www.mitsuba-renderer.org/
• http://www.pbrt.org/
• http://www.luxrender.net/en_GB/index

32
Questions?
• Vintage path tracing by Kajiya

© Jim Kajiya. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

33
Path Tracing is costly
• Needs tons of rays per pixel!

34
Global Illumination (with Indirect)

Courtesy of Henrik Wann Jensen. Used with permission.


35
Indirect Lighting is Mostly Smooth

Courtesy of Henrik Wann Jensen. Used with permission.


36
Irradiance Caching
• Indirect illumination is smooth

37
Irradiance Caching
• Indirect illumination is smooth

38
Irradiance Caching
• Indirect illumination is smooth
==> Sample sparsely, interpolate nearby values

39
Irradiance Caching
• Store the indirect illumination
• Interpolate existing cached values
• But do full calculation for direct lighting

40
Irradiance Caching
• Yellow dots:
indirect diffuse sample points

The irradiance cache tries to


adapt sampling density to
expected frequency content of
the indirect illumination (denser
sampling near geometry)
Courtesy of Henrik Wann Jensen. Used with permission.
41
Radiance by Greg Ward
• The inventor of irradiance caching
• http://radsite.lbl.gov/radiance/

Image removed due to copyright restrictions. Please see above link for further details.

42
Questions?

Image of Y chair designed by H.J. Wegner has been removed due to copyright restrictions.
Please see http://tora_2097.cgsociety.org/portfolio/project-detail/786738/ for further details.

Image: Pure
43
Photon Mapping
• Preprocess: cast rays from light sources, let them
bounce around randomly in the scene
• Store “photons”

44
Photon Mapping
• Preprocess: cast rays from light sources
• Store photons (position + light power + incoming direction)

45
The Photon Map
• Efficiently store photons for fast access
• Use hierarchical spatial structure (kd-tree)

46
Photon Mapping - Rendering
• Cast primary rays
• For secondary rays
– reconstruct irradiance using adjacent stored photon
– Take the k closest photons
• Combine with irradiance caching and a number of other techniques

Shooting one bounce of


secondary rays and
using the density
approximation at those
hit points is called final
gathering.

47
Photon Map Results

Courtesy of Henrik Wann Jensen. Used with permission.

48
More Global Illumination Coolness
• Many materials exhibit subsurface scattering
– Light doesn’t just reflect off the surface
– Light enters, scatters around, and exits at another point
– Examples: Skin, marble, milk
Images: Jensen et al.

Courtesy of Henrik Wann Jensen. Used with permission. 49


More Subsurface Scattering
Weyrich et al. 2006

Photograph Rendering
© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
50
That Was Just the Beginning
• Tons and tons of other Monte Carlo techniques
– Bidirectional Path Tracing
• Shoot random paths not just from camera but also light, connect
the path vertices by shadow rays
– Metropolis Light Transport
• And Finite Element Methods
– Use basis functions instead of random sampling
– Radiosity (with hierarchies & wavelets)
– Precomputed Radiance Transfer

• This would warrant a class of its own!


51
What Else Can We Integrate?
• Pixel: antialiasing
• Light sources: Soft shadows
• Lens: Depth of field
• Time: Motion blur
• BRDF: glossy reflection
• (Hemisphere: indirect lighting) Courtesy of Henrik Wann Jensen.
Used with permission.

© ACM. All rights reserved. This content is © source unknown. All rights reserved.
© source unknown. All rights reserved. This content is excluded from our Creative Commons This content is excluded from our Creative
excluded from our Creative Commons license. For more license. For more information, see Commons license. For more information,
information, see http://ocw.mit.edu/help/faq-fair-use/. http://ocw.mit.edu/help/faq-fair-use/. see http://ocw.mit.edu/help/faq-fair-use/. 52
Domains of Integration
• Pixel, lens (Euclidean 2D domain)
– Antialiasing filters, depth of field
• Time (1D) Famous motion blur image
from Cook et al. 1984
– Motion blur
• Hemisphere
– Indirect lighting
• Light source
– Soft shadows

© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
53
Motivational Eye Candy
• Rendering glossy reflections
• Random reflection rays around mirror direction
– 1 sample per pixel

© source unknown. All rights reserved. This content is excluded from our Creative
Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
54
Motivational Eye Candy
• Rendering glossy reflections
• Random reflection rays around mirror direction
– 256 samples per pixel

© source unknown. All rights reserved. This content is excluded from our Creative
Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
55
Error/noise Results in Variance
• We use random rays
– Run the algorithm again  get different image
• What is the noise/variance/standard deviation?
– And what’s really going on anyway?

© source unknown. All rights reserved. This content is excluded from our Creative
Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/. 56
Integration
• Compute integral of arbitrary function
– e.g. integral over area light source, over hemisphere, etc.
• Continuous problem  we need to discretize
– Analytic integration never works because of visibility and other
nasty details

57
Integration
• You know trapezoid, Simpson’s rule, etc.

58
Monte Carlo Integration
• Monte Carlo integration: use random samples and
compute average
– We don’t keep track of spacing between samples
– But we kind of hope it will be 1/N on average

59
Monte Carlo Integration

• S is the integration domain


– Vol(S) is the volume (measure) of S
• {xi} are independent uniform random points in S

60
Monte Carlo Integration

• S is the integration domain


– Vol(S) is the volume (measure) of S
• {xi} are independent uniform random points in S
• The integral is the average of f times the volume of S
• Variance is proportional to 1/N
– Avg. error is proportional 1/sqrt(N)
– To halve error, need 4x samples
61
Monte Carlo Computation of 
• Take a square
• Take a random point (x,y) in the square
• Test if it is inside the ¼ disc (x2+y2 < 1)
• The probability is  /4

y Integral of the function that


is one inside the circle, zero
outside
x

62
Monte Carlo Computation of 
• The probability is  /4
• Count the inside ratio n = # inside / total # trials
• n*4
• The error depends on the number or trials
Demo
def piMC(n):
success = 0
for i in range(n): x=random.random()
y=random.random()
if x*x+y*y<1: success = success+1 return
4.0*float(success)/float(n)

63
Why Not Use Simpson Integration?
• You’re right, Monte Carlo is not very efficient for
computing 
• When is it useful?
– High dimensions: Convergence is independent of
dimension!
– For d dimensions, Simpson requires Nd domains (!!!)
– Similar explosion for other quadratures (Gaussian, etc.)

64
Advantages of MC Integration
• Few restrictions on the integrand
– Doesn’t need to be continuous, smooth, ...
– Only need to be able to evaluate at a point
• Extends to high-dimensional problems
– Same convergence
• Conceptually straightforward
• Efficient for solving at just a few points

65
Disadvantages of MC
• Noisy
• Slow convergence
• Good implementation is hard
– Debugging code
– Debugging math
– Choosing appropriate techniques

66
Questions?
• Images by Veach and Guibas, SIGGRAPH 95

Naïve sampling strategy Optimal sampling strategy


© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
67
Hmmh...
• Are uniform samples the best we can do?

68
Smarter Sampling
Sample a non-uniform probability
Called “importance sampling”
Intuitive justification: Sample more in places where there are
likely to be larger contributions to the integral

69
Example: Glossy Reflection
Slide courtesy of Jason Lawrence
• Integral over hemisphere
• BRDF times cosine times incoming light

Image removed due to copyright restrictions – please see Jason Lawrence’s slide 9-12 in the talk slides on “Efficient BRDF
Importance Sampling Using a Factored Representation,” available at http://www.cs.virginia.edu/~jdl/.

70
Sampling a BRDF Slide courtesy of Jason Lawrence

Image removed due to copyright restrictions – please see Jason Lawrence’s slide 9-12 in the talk slides on “Efficient BRDF
Importance Sampling Using a Factored Representation,” available at http://www.cs.virginia.edu/~jdl/.

71
Sampling a BRDF Slide courtesy of Jason Lawrence

Image removed due to copyright restrictions – please see Jason Lawrence’s slide 9-12 in the talk slides on “Efficient BRDF
Importance Sampling Using a Factored Representation,” available at http://www.cs.virginia.edu/~jdl/.

72
Sampling a BRDF Slide courtesy of Jason Lawrence

Image removed due to copyright restrictions – please see Jason Lawrence’s slide 9-12 in the talk slides on “Efficient BRDF
Importance Sampling Using a Factored Representation,” available at http://www.cs.virginia.edu/~jdl/.

73
Importance Sampling Math

• Like before, but now {xi} are not uniform but drawn
according to a probability distribution p
– Uniform case reduces to this with p(x) = const.
• The problem is designing ps that are easy to sample
from and mimic the behavior of f

74
Monte Carlo Path Tracing

Video removed due to copyright restrictions – please see the link below for further details.

http://www.youtube.com/watch?v=mYMkAnm-PWw 75
Questions?
1200 Samples/Pixel

Traditional importance function Better importance by Lawrence et al.


© ACM. All rights reserved. This content is excluded from our Creative Commons
license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.
76
Stratified Sampling
• With uniform sampling, we can get unlucky
– E.g. all samples clump in a corner
– If we don’t know anything of the integrand,
we want a relatively uniform sampling
• Not regular, though, because of aliasing!

• To prevent clumping, subdivide domain 


into non-overlapping regions i
– Each region is called a stratum
• Take one random sample per i

77
Stratified Sampling Example
• When supersampling, instead of taking KxK regular
sub-pixel samples, do random jittering within each
KxK sub-pixel

78
Stratified Sampling Analysis
• Cheap and effective
• But mostly for low-dimensional domains
– Again, subdivision of N-D needs Nd domains like
trapezoid, Simpson’s, etc.!

• With very high dimensions, Monte Carlo is pretty


much the only choice

79
Questions?
• Image from the ARNOLD Renderer by Marcos Fajardo

Images removed due to copyright restrictions -- Please see


http://www.3dluvr.com/marcosss/morearni/ for further details.

80
For Further Information...
• 6.839!
• Eric Veach’s PhD dissertation
http://graphics.stanford.edu/papers/veach_thesis/

• Physically Based Rendering


by Matt Pharr, Greg Humphreys

81
References

Images of the following book covers have been removed due to copyright restrictions:
-Advanced Global Illumination by Philip Dutre, Philippe Bekaert, and Kavita Bala
-Realistic Ray Tracing by Peter Shirley and R. K. Morley
-Realistic Image Synthesis Using Photon Mapping by Henrik Wann Jensen
Please check the books for further details.

82
That’s All for today

Image removed due to copyright restrictions -- please Fig. 13 in Fournier A. and W.T. Reeves. "A Simple Model of Ocean Waves."
SIGGRAPH '86 Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques; Pages 75-84.

Image: Fournier and


Reeves, SIGGRAPH 86 83
MIT OpenCourseWare
http://ocw.mit.edu

6.837 Computer Graphics


Fall 2012

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

You might also like