CS 431/636

Advanced Rendering
Techniques

Dr. David Breen

Matheson 308
Thursday 6PM 
8:50PM
Presentation 5
5/9/06
Questions from Last
Time?
Acceleration techniques
 Bounding Regions
 Acceleration Spatial Data
Structures
 Regular Grids
 Adaptive Grids

 Bounding Volume Hierarchies

 Dot product problem

Slide Credits

 David Luebke - University of

Virginia
 Kevin Suffern - University of
Technology, Sydney, Australia
Anti-aliasing
 Aliasing: signal processing
term with very specific meaning
 Aliasing: computer graphics
term for any unwanted visual
artifact based on sampling
 Jaggies, drop-outs, etc.
 Anti-aliasing: computer
graphics term for fixing these
unwanted artifacts
 We’ll tackle these in order
Signal Processing
 Raster display: regular
sampling of a continuous(?)
function
 Think about sampling a 1-D
function:
Signal Processing
 Sampling a 1-D function:
Signal Processing
 Sampling a 1-D function:
Signal Processing
 Sampling a 1-D function:
 What do you notice?
Signal Processing
 Sampling a 1-D function: what
do you notice?
 Jagged, not smooth
Signal Processing
 Sampling a 1-D function: what
do you notice?
 Jagged, not smooth
 Loses information!
Signal Processing
 Sampling a 1-D function: what
do you notice?
 Jagged, not smooth
 Loses information!

 What can we do about these?

 Use higher-order reconstruction
 Use more samples
 How many more samples?
The Sampling Theorem
 Obviously, the more samples we
take the better those samples
approximate the original
function
 The sampling theorem:
A continuous bandlimited function can be
completely represented by a set of equally
spaced samples, if the samples occur at
more than twice the frequency of the
highest frequency component of the function
The Sampling Theorem
 In other words, to adequately
capture a function with maximum
frequency F, we need to sample
it at frequency N = 2F.
 N is called the Nyquist limit
 What is the maximum frequency?
The Sampling Theorem
 An example: sinusoids
The Sampling Theorem
 An example: sinusoids
Fourier Theory
 Fourier theory lets us
decompose any signal into the
sum of (a possibly infinite
number of) sine waves
 Next comes the math…
 Switch to hand-written notes
Recap: Fourier Theory
 Can decompose any signal I(x) into
a spectrum of sine waves F(u)
 F(u) is complex; we represent it
with:
|F(u)| = [Real(u)2 + Imag(u)2]1/2
Ø(u) = tan-1[Real(u)/Imag(u)]
 These are called the amplitude and
phase spectra
Recap: Aliasing
 Sampling a function
= multiplying I(x) by sampling fn Sx
= convolving F(u) by sampling fn S1/x
= copying F(u) at regular intervals
 If the copies of F(u) overlap, high
frequencies “fold over”, appearing
as lower frequencies
 This is known as aliasing
Recap: Prefiltering
 Eliminate high frequencies
before sampling
 Convert I(x) to F(u)
 Apply a low-pass filter (e.g.,
multiply F(u) by a box function)
 Then sample. Result: no aliasing!
 Problem: most rendering
algorithms generate sampled
function directly
 e.g., Z-buffer, ray tracing
Supersampling
 The simplest way to reduce aliasing
artifacts is supersampling
 Increase the resolution of the samples
 Average the results down
 Sometimes called postfiltering
 Create virtual image at higher
resolution than the final image
 Apply a low-pass filter
 Resample filtered image
Supersampling:
Limitations
 Q: What practical consideration
hampers supersampling?
 A: Storage goes up quadratically
 Q: What theoretical problem does
supersampling suffer?
 A: Doesn’t eliminate aliasing!
Supersampling simply shifts the
Nyquist limit higher
 Supersampling:
Worst Case

 Q: Give a
simple scene
containing
infinite
frequencies
 A: A checkered
ground plane
receding into
the background
Supersampling
 Create virtual image at higher
resolution than the final
image
 This is easy
Supersampling
 Apply a low-pass filter
 Convert to frequency domain
 Multiply by a box function

 Expensive! Can we simplify this?

 Recall: multiplication in frequency
equals convolution in space, so…
 Just convolve initial sampled image
with the FT of a box function
 Isn’t this a lot of work?
Supersampling
 Q: What filter should we use in
space to effect multiplying by a
box function in frequency?
 A: The sinc function (sin(x) / x)
 Q: Why is this hard?
 Sinc function has infinite support
 Sinc function has negative lobes
Supersampling
 Other common filters:
 Truncated sinc
 Box

 Triangle

 Gaussian

 Each has their own frequency

spectrum and artifacts
Supersampling:
Digital Convolution
 In practice, we combine steps 2 &
3:
 Create virtual image at higher
resolution than the final image
 Apply a low-pass filter
 Resample filtered image

 Idea: only create filtered image

at new sample points
 I.e., only convolve filter with
image at new points
Supersampling:
Digital Convolution
 Q: What does convolving a
filter with an image entail
at each sample point?
 A: Multiplying and summing 1 2 1
values = 2 4 2
1 2 1
 Example (also F&vD p 642): 1 2 1

 =
1 2 1 2 4 2
2 4 2 1 2 1
1 2 1

…
Supersampling
 Typical supersampling algorithm:
 Compute multiple samples per pixel
 Combine sample values for pixel’s value
using simple average
 Q: What filter does this equate to?
 A: Box filter -- one of the worst!
 Q: What’s wrong with box filters?
 Passes infinitely high frequencies
 Attenuates desired frequencies
Supersampling
In Practice
 Sinc function: ideal but
impractical
 One approximation: sinc2
 Another: Guassian falloff
 Q: How wide (what res) should
filter be?
 A: As wide as possible!
In practice: 3x3,
5x5, at most 7x7
Supersampling:
Summary
 Supersampling improves aliasing
artifacts by shifting the Nyquist
limit
 It works by calculating a high-res
image and filtering down to final res
 “Filtering down” means simultaneous
convolution and resampling
 This equates to a weighted average
 Wider filterbetter resultsmore work
Supersampling:
Cons & Pros
 Supersampling cons
 Doesn’t eliminate aliasing, just shifts
the Nyquist limit higher
• Can’t fix some scenes (e.g., checkerboard)
 Badly inflates storage requirements
 Supersampling pros
 Relatively easy
 Often works all right in practice
 Can be added to a standard renderer
 Catmull’s Algorithm

A2
 Find fragment
areas
A1
 Multiply by
fragment
colors
 Sum for final
pixel color
AB A3
Catmull’s Algorithm
 First real attempt to filter
in continuous domain
 Very expensive
 Clipping polygons to fragments
 Sorting polygon fragments by

depth
 Equates to box filter
The A-Buffer
 Idea: approximate continuous filtering
by subpixel sampling
 Summing areas now becomes simple
The A-Buffer
 Advantages:
 Incorporating into scanline
renderer reduces storage costs
dramatically
 Processing per pixel depends only

on number of visible fragments

 Can be implemented efficiently

using bitwise logical ops on

subpixel masks
The A-Buffer
 Disadvantages
 Still basically a supersampling
algorithm
 Not a hardware-friendly algorithm

• Lists of potentially visible polygons

can grow without limit
• Work per-pixel non-deterministic
Stochastic Sampling
 Sampling theory tells us that
with a regular sampling grid,
frequencies higher than the
Nyquist limit will alias
 Q: What about irregular sampling?
 A: High frequencies appear as
noise, not aliases
 This turns out to bother our
visual system less!
Stochastic Sampling
 An intuitive argument:
 In stochastic sampling, every
region of the image has a finite
probability of being sampled
 Thus small features that fall

between uniform sample points

tend to be detected by non-
uniform samples
Stochastic Sampling
 Idea: randomizing distribution of
samples scatters aliases into
noise
 Problem: what type of random
distribution to adopt?
 Reason: type of randomness used
affects spectral characteristics
of noise into which high
frequencies are converted
Stochastic Sampling
 Problem: given a pixel, how to
distribute points (samples)
within it?
Stochastic Sampling
 Poisson distribution:
 Completely random
 Add points at random until area

is full.
 Uniform distribution: some

neighboring samples close

together, some distant
Stochastic Sampling
 Poisson disc distribution:
 Poisson distribution, with
minimum-distance constraint
between samples
 Add points at random, removing

again if they are too close to

any previous points
 Very even-looking distribution
Stochastic Sampling
 Jittered distribution
 Start with regular grid of
samples
 Perturb each sample slightly in a

random direction
 More “clumpy” or granular in

appearance
Stochastic Sampling
 Spectral characteristics of these
distributions:
 Poisson: completely uniform (white noise).
High and low frequencies equally present
 Poisson disc: Pulse at origin (DC
component of image), surrounded by empty
ring (no low frequencies), surrounded by
white noise
 Jitter: Approximates Poisson disc
spectrum, but with a smaller empty disc.
 Comparison

Regular Random

Poisson Jitter
disk
Recap:
Antialiasing Strategies
 Prefiltering: low-pass filter
the signal before sampling
 Pros:
• Guaranteed to eliminate aliasing
• Preserves all desired frequencies
 Cons:
• Expensive
• Can introduce “ringing”
• Doesn’t fit most rendering algorithms
Recap:
Antialiasing Strategies
 Supersampling:sample at
higher resolution, then
filter down
 Pros:
• Conceptually simple
• Easy to retrofit existing renderers
• Works well most of the time
 Cons:
• High storage costs
• Doesn’t eliminate aliasing, just
shifts Nyquist limit upwards
Recap:
Antialiasing Strategies
 A-Buffer: approximate
prefiltering of continuous signal
by sampling
 Pros:
• Integrating with scan-line renderer
keeps storage costs low
• Can be efficiently implemented with
clever bitwise operations
 Cons:
• Still basically a supersampling approach
• Doesn’t integrate with ray-tracing
Recap:
Antialiasing Strategies
 Stochastic supersampling:
convert high frequencies into
noise
 Pros:
• Retains advantages of supersampling
• Noise is visually better than
aliasing
 Cons:
• Ideal sampling distribution (Poisson
Disc) is expensive
Go to Suffern Slides
Wrap Up
 Discuss next programming
assignment
 Add adaptive supersampling
 Discuss
status/problems/issues with
this week’s programming
assignment