Advanced Computer Graphics

Aliasing
Matthias Teschner
Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 2

Motivation
− Sampling and
reconstruction
− Inappropriate
sampling can
cause artifacts
in reconstructed
functions

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 3
Motivation
− Aliasing artifacts, e. g. Moiré pattern

Original signal Sampled signal Reconstructed signal

Red - original signal

Black dots - samples
[Wikipedia:
Blue - reconstructed signal Alias-Effekt]
University of Freiburg – Computer Science Department – 4
Motivation
− Aliasing artifacts, e. g. jaggies

[Suffern]

University of Freiburg – Computer Science Department – 5

Motivation
− Aliasing artifacts, e. g. missing details

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 6

Motivation
− Antialiasing
− Reduction of erroneous patterns

Regular artifacts replaced

Aliasing artifacts [Suffern]
by less disturbing noise

University of Freiburg – Computer Science Department – 7

Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 8

Spectrum of a Function
− Fourier transform
− Decomposes a function into
weighted sum of shifted sinusoids
− Computes amplitude and
phase shift of frequencies
contained in the function
− Transforms from the spatial
domain to the frequency domain
[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 9

Spectrum of a Function - Motivation
− Analysis in the frequency
Appropriate
domain allows to sampling
understand aliasing frequencies

− Aliasing is then reduced by

− Adapting the sampling
− Filtering the original signal Inappropriate
sampling
(for textures) frequencies

[Wikipedia: Nyquist-Shannon-Abtasttheorem]

University of Freiburg – Computer Science Department – 10

Spectrum of a Function - Motivation
− Sampling and reconstruction can be analyzed
in the frequency domain
− A band-limited function with for all
has to be sampled with a frequency in
order to be able to reconstruct the
original function from the samples
(Nyquist-Shannon sampling theorem)
− Nyquist frequency
− Nyquist rate
University of Freiburg – Computer Science Department – 11
Fourier Transform - Examples
− Signals in spatial and
frequency domain

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 12
Fourier Transform - Properties
− Fourier transform of the product of two functions
is equivalent to the convolution of the individual
Fourier transforms

− Convolution in the spatial domain is equivalent to

multiplication in the frequency domain

− Important in understanding how filtering and

reconstruction affect the spectrum of a function
University of Freiburg – Computer Science Department – 13
Convolution

− Convolution computes a weighted average

of f using the weighting kernel g

[Wikipedia: Convolution]

University of Freiburg – Computer Science Department – 14

Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 15

Low-Pass Filtering
− Convolution is used to filter and reconstruct functions
Original function

sinc function

Band-limited signal

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 16
Low-Pass Filtering
− sinc function in spatial and frequency domain

sinc function Box function

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 17

Low-Pass Filtering
− Convolution with sinc function in the spatial
domain corresponds to multiplication with
a box function in the frequency domain
− Given a sampling rate, this low-pass filter
completely suppresses all frequency
components above the Nyquist frequency
− Aliasing is avoided in the reconstruction process
− Applied in texturing
− In ray tracing, the original function
cannot be simply filtered [Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 18

Approximate Low-Pass Filtering
− Truncated sinc
− Gibbs phenomenon

− Triangle

− Gaussian

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 19
Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 20

Sampling
− Sampling a function corresponds to multiplying
it in the spatial domain by a Dirac comb function

Original function Dirac comb function Sampled function

(shah function)
[Foley, van Dam, Feiner, Huges]
University of Freiburg – Computer Science Department – 21
Sampling
− In frequency domain, sampling is a convolution of
the function's spectrum with a Dirac comb function
Dirac comb in the Dirac comb in the
spatial domain frequency domain

Spectrum of Spectrum of
the function the sampled
function

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 22
Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 23

Reconstruction
− As a result of sampling, F contains an infinite number
of replications of the spectrum of the original function f
at multiples of the sampling frequency
− Reconstruction tries to remove all but the spectrum
of the original function by multiplying with a box filter
in the frequency domain (corresponding to a convolution
of the sampled function with sinc in the spatial domain)
− Our visual system reconstructs a function from pixel values
− In ray tracing, incident radiance at pixels can be reconstructed
from several samples
University of Freiburg – Computer Science Department – 24
Reconstruction
− Sampling and
reconstruction
in spatial and
frequency
domain

[Foley, van Dam, Feiner, Huges]

University of Freiburg – Computer Science Department – 25
Aliasing
− If the sampling frequency is too low, the replicated copies
of the spectra overlap and the spectrum of the original
function cannot be reconstructed
Sufficient
sampling

Non-
overlapping
spectra
Reconstruction

Insufficient
sampling
[Foley, van Dam,
Feiner, Huges]
Overlapping
Original function in case of sufficient sampling
spectra
University of Freiburg – Computer Science Department – 26
Aliasing

Spectrum of a band- [Wikipedia: Nyquist-Shannon-Abtasttheorem]

limited function

Spectrum of sufficiently sampled Spectrum of an insufficiently sampled

band-limited function, a box filter band-limited function, the original spectrum
can reconstruct the original spectrum cannot be reconstructed

University of Freiburg – Computer Science Department – 27

Outline
− Motivation
− Fourier analysis
− Filtering
− Sampling
− Reconstruction / aliasing
− Antialiasing

University of Freiburg – Computer Science Department – 28

Antialiasing
− In texturing, textures are filtered according
to the given sampling rate
− prefiltering
− In ray tracing, the sampling rate and
sampling patterns are adapted
− Nonuniform sampling: tends to turn regular aliasing patterns into noise
− Adaptive sampling: use more samples in case of large variations bet-
ween adjacent samples (might still miss high frequencies, small details)
− In ray tracing, radiance at a pixel position is commonly
reconstructed from samples within the pixel area and
samples in adjacent pixels
University of Freiburg – Computer Science Department – 29
Antialiasing

Regular sampling Non-uniform, random

[Suffern]
with aliasing sampling with noise

University of Freiburg – Computer Science Department – 30

Summary
− Sampling of the continuous radiance function can cause aliasing
− Moiré patterns
− Jaggies
− Missing details
− Fourier analysis helps to understand sampling,
filtering / reconstruction, and aliasing effects
− Fourier transform converts between
spatial and frequency domain
− Fourier transform of the product of two functions is equivalent
to the convolution of the individual Fourier transforms
University of Freiburg – Computer Science Department – 31