U.S. patent application number 12/246273 was filed with the patent office on 2009-06-18 for image synthesis methods and systems.
Invention is credited to Alexander Keller.
Application Number | 20090153576 12/246273 |
Document ID | / |
Family ID | 46325918 |
Filed Date | 2009-06-18 |
United States Patent
Application |
20090153576 |
Kind Code |
A1 |
Keller; Alexander |
June 18, 2009 |
Image Synthesis Methods and Systems
Abstract
The present invention provides systems, devices,
computer-implemented methods and computer program code products
(software) operable to evaluate integrals using quasi-Monte Carlo
methodologies, and in particular embodiments, adaptive quasi-Monte
Carlo integration and adaptive integro-approximation in conjunction
with techniques including a scrambled Halton Sequence,
stratification by radical inversion, stratified samples from the
Halton Sequence, deterministic scrambling, bias elimination by
randomization, adaptive and deterministic anti-aliasing,
anti-aliasing by rank-1 lattices, and trajectory splitting by
dependent sampling and rank-1 lattices.
Inventors: |
Keller; Alexander; (Ulm,
DE) |
Correspondence
Address: |
JACOBS & KIM LLP
1050 WINTER STREET, SUITE 1000, #1082
WALTHAM
MA
02451-1401
US
|
Family ID: |
46325918 |
Appl. No.: |
12/246273 |
Filed: |
October 6, 2008 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
11465717 |
Aug 18, 2006 |
7432935 |
|
|
12246273 |
|
|
|
|
10299958 |
Nov 19, 2002 |
7167175 |
|
|
11465717 |
|
|
|
|
09884861 |
Jun 19, 2001 |
7227547 |
|
|
10299958 |
|
|
|
|
60709173 |
Aug 18, 2005 |
|
|
|
60265934 |
Feb 1, 2001 |
|
|
|
60212286 |
Jun 19, 2000 |
|
|
|
Current U.S.
Class: |
345/581 |
Current CPC
Class: |
G06T 15/06 20130101;
G06T 15/506 20130101; G06T 15/55 20130101; G06T 11/001 20130101;
G06F 17/10 20130101 |
Class at
Publication: |
345/581 |
International
Class: |
G09G 5/00 20060101
G09G005/00 |
Claims
1. A computer-implemented method of generating a pixel value for a
pixel in an image displayable via a display device, the pixel value
being representative of a point in a scene, the method comprising:
A. generating a set of sample points, at least one sample point
being generated using at least one sample, the at least one sample
comprising at least one element of a sequence, and wherein the set
of sample points comprises quasi-Monte Carlo points; and B.
evaluating a selected function at one of the sample points to
generate a value, the generated value corresponding to the pixel
value, the pixel value being usable to generate a
display-controlling electronic output.
2. The method of claim 1 wherein the sequence comprises a scrambled
Halton sequence.
3. The method of claim 1 wherein, in the sequence, a radical
inverse function is replaced with a scrambled radical inverse
function to yield a scrambled Halton sequence.
4. The method of claim 3 wherein the replacement of the radical
inverse with the scrambled radical inverse is described by the
following equation: .PHI. b ' : 0 .fwdarw. [ 0 , 1 ) ##EQU00085## i
= l = 0 .infin. a l ( i ) b l l = 0 .infin. .pi. b ( a l ( i ) ) b
- l - 1 , ##EQU00085.2## wherein .pi..sub.b is a scrambling
permutation applied to the digits a.sub.g(i), and wherein the
scrambling permutation is determined by a permutation of the set of
integers {0, . . . , b-1}.
5. A computer-implemented method of generating a pixel value for a
pixel in an image display able via a display device, the pixel
value being representative of a point in a scene, the method
comprising: A. generating a set of sample points, at least one
sample point being generated using at least one sample, the at
least one sample comprising at least one element of a
low-discrepancy sequence, and wherein the generating includes using
an adaptive, interleaved sampling scheme based on a
deterministically scrambled Halton sequence to yield a
deterministic, low-discrepancy set of sample points; and B.
evaluating a selected function at one of the sample points to
generate a value, the generated value corresponding to the pixel
value, the pixel value being usable to generate a
display-controlling electronic output.
6. A computer-implemented method of generating a pixel value for a
pixel in an image displayable via a display device, the pixel value
being representative of a point in a scene, the method comprising:
A. generating a set of sample points, at least one sample point
being generated using at least one sample, the at least one sample
comprising at least one element of a sequence, wherein the set of
sample points comprises quasi-Monte Carlo points, and wherein the
generating includes adaptively sampling by using radical
inversion-based points; and B. evaluating a selected function at
one of the sample points to generate a value, the generated value
corresponding to the pixel value, the pixel value being usable to
generate a display-controlling electronic output.
7. The method of claim 6 wherein the sequence is a Halton
sequence.
8. The method of claim 6 wherein the sequence is a (t, s)
sequence.
9. The method of claim 6 wherein the generating includes
constructing multi-dimensional, substantially uniform deterministic
samples using a Halton sequence.
10. The method of claim 6 wherein the generating includes using
sets of points having maximized minimum distance.
11. The method of claim 6 wherein the generating includes using
point sets having maximized minimum distance with respect to low
dimensional projections.
12. The method of claim 6 wherein the sequence is a scrambled
Halton sequence, and wherein the generating further comprises
extracting stratified sequences of sample points from the scrambled
Halton sequence.
13. The method of claim 6 wherein the sequence is a (t, s)
sequence, and wherein the generating further comprises extracting
stratified sequences of sample points from the (t, s) sequence.
14. The method of claim 6 wherein the evaluating comprises
evaluating a pixel integral, and further comprising applying a tone
mapping function within the pixel integral, so as to improve
convergence.
15. The method of claim 14 wherein applying a tone mapping function
comprises applying a tone-mapping function that bounds the
integrands, so as to improve convergence.
16. The method of claim 14 further comprising controlling
adaptation by applying image processing operators.
17. The method of claim 14 further comprising controlling
adaptation by applying image processing operators to final pixel
values rather than to single samples.
18. The method of claim 12 further comprising providing bias
elimination by randomization, wherein the randomization comprises
any of scrambling, or randomly shifting deterministic points of the
scrambled Halton sequence modulo one.
19. The method of claim 7 further comprising providing
deterministic anti-aliasing by scaling a first component of a
sampling function by a first scaling coefficient, and a second
component by a second scaling coefficient, to obtain a stratified
sample pattern that can be periodically tiled over an image plane,
the stratified sample pattern having a plurality of strata;
identifying each stratum with a given pixel; and after identifying
each stratum with a given pixel, obtaining a per-stratum
identification value from the pixel coordinates and generating a
Halton sequence specific to a corresponding sample based on the
corresponding identification value.
20. The method of claim 7 further comprising providing adaptive
anti-aliasing by stratification by the Halton sequence.
21. The method of claim 19 further comprising determining the
number of strata by selecting exponents, for the first and second
scaling coefficients, large enough so that strata covered by
adjacent pixel reconstruction filters do not contain repeated
patterns.
22. The method of claim 6 wherein a pixel is deemed refined
whenever a refinement criterion is met, and wherein the refinement
criterion can include comparing the image gradient against a
predefined threshold T.
23. The method of claim 22 further comprising selecting a value,
for an exponent for a coefficient to be multiplicatively applied to
the threshold T, to enable adaptation to the speed of convergence,
wherein the coefficient is the sampling rate.
24. The method of claim 14 wherein the tone mapping comprises
compression of a luminance value L prior to averaging, and wherein
the luminance value L is compressed in accordance with the
following compression equation: R .alpha. : 0 + .fwdarw. [ 0 , 1 ]
##EQU00086## L { L L < .alpha. .alpha. + ( 1 - .alpha. ) L -
.alpha. 1 + L - .alpha. else ##EQU00086.2## wherein .alpha. is a
coefficient selected to be between 0 and 1, and R.sub..alpha. is a
response curve mapping that can be selected by selection of
coefficient .alpha..
25. A computer-implemented method of generating a pixel value for a
pixel in an image displayable via a display device, the pixel value
being representative of a point in a scene, the method comprising:
A. generating a set of sample points, at least one sample point
being generated using at least one sample, the at least one sample
comprising at least one element of a sequence, wherein the
generating includes sampling by using rank-1 lattice points; and B.
evaluating a selected function at one of the sample points to
generate a value, the generated value corresponding to the pixel
value, the pixel value being usable to generate a
display-controlling electronic output.
26. The method of claim 25 further comprising selecting a rank-1
lattice such that its mutual minimum distance among sample points
is maximal.
27. The method of claim 25 further wherein the generating includes
using lattice sequences.
28. The method of claim 25 wherein the lattices are lattices in
Korobov form.
29. The method of claim 25 wherein the lattices are rank-1 lattices
or higher rank lattices in Korobov or general form.
30. The method of claim 25 further comprising using lattices with
respect to low-dimensional projections of the points.
31. The method of claim 25 further comprising applying
anti-aliasing by lattices, including adding to the lattice points a
different random shift to generate a randomly shifted lattice,
thereby to attenuate aliasing over the pixels.
32. The method of claim 31 further comprising de-randomizing the
random shifts per pixel by determining a shift per pixel by
elements of a low discrepancy point set or a deterministic point
set with maximized minimum distance, with stratification that
matches the pixels.
33. The method of claim 32 wherein the stratification is induced by
a rank-1 lattice.
34. The method of claim 32 wherein the stratification is induced by
the Voronoi diagram of a rank-1 lattice.
35. The method of claim 32 further comprising using recursive
Korobov filters wherein the points inside a given lattice cell are
determined by another set of lattice points transformed into the
given lattice cell.
36. The method of claim 25 further comprising applying trajectory
splitting using domain stratification induced by a rank-1 lattice
with maximized minimum distance.
37. The method of claim 25 further comprising providing quasi-Monte
Carlo integro-approximation by lattice points, and wherein point
sets and sequences are selected by maximum minimum distance.
38. In a computer graphics system including a processor, a display
device, user input elements, and one or more memory elements, the
computer graphics system being operable to generate images
displayable via a display device, the images representing a scene
and comprising a plurality of pixels, a computer-implemented system
for generating a pixel value for a pixel in an image displayable
via the display device, the pixel value being representative of a
point in a scene, the system comprising: A. means for generating a
set of sample points, at least one sample point being generated
using at least one sample, the at least one sample comprising at
least one element of a sequence, and wherein the set of sample
points comprises quasi-Monte Carlo points; and B. means, in
communication with the means for generating a set of sample points,
for evaluating a selected function at one of the sample points to
generate a value, the generated value corresponding to the pixel
value, the pixel value being usable to generate a
display-controlling electronic output.
39. A computer program product for use in a computer graphics
system, for enabling the computer graphics system to generate a
pixel value for a pixel in an image displayable via a display
device, the pixel value being representative of a point in a scene,
the computer program product comprising a computer-readable medium
having encoded thereon: A. computer-readable program instructions
executable to enable the computer graphics system to generate a set
of sample points, at least one sample point being generated using
at least one sample, the at least one sample comprising at least
one element of a sequence, and wherein the set of sample points
comprises quasi-Monte Carlo points; and B. computer-readable
program instructions executable to enable the computer graphics
system to evaluate a selected function at one of the sample points
to generate a value, the generated value corresponding to the pixel
value, the pixel value being usable to generate a
display-controlling electronic output.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS AND INCORPORATION BY
REFERENCE
[0001] This application for patent is a Continuation of U.S. patent
application Ser. No. 11/465,717 filed Aug. 18, 2006 (Atty. Dkt.
MENT-104-US). Patent application Ser. No. 11/465,717 claims the
priority benefit of U.S. Provisional Patent App. 60/709,173 filed
Aug. 18, 2005, and is a continuation-in-part of U.S. patent
application Ser. No. 10/299,958 filed Nov. 19, 2002 (which issued
as U.S. Pat. No. 7,167,175 on Jan. 23, 2007) (MENT-072). U.S.
patent application Ser. No. 10/299,958 is a Continuation-in-Part of
U.S. patent application Ser. No. 09/884,861 filed Jun. 19, 2001
(which issued as U.S. Pat. No. 7,227,547 on Jun. 5, 2007)
(MENT-061), which claims priority from U.S. Provisional Patent
Apps. 60/265,934 filed Feb. 1, 2001 and 60/212,286 filed Jun. 19,
2000 (both expired). Each of the patent applications noted above is
incorporated herein by reference as if set forth herein in its
entirety.
[0002] Also incorporated herein by reference is commonly owned U.S.
patent application Ser. No. 08/880,418, filed Jun. 23, 1997, in the
names of Rolf Herken and Martin Grabenstein (Attorney Docket
MENT-002), now U.S. Pat. No. 6,529,193, entitled "System and Method
for Generating Pixel Values for Pixels in an Image Using Strictly
Deterministic Methodologies for Generating Sample Points,"
hereinafter referred to as "Grabenstein."
FIELD OF THE INVENTION
[0003] Hie present invention relates generally to methods, systems
and computer program code products (software) for image synthesis
in and by digital computing systems, such as for motion pictures
and other computer graphics applications, and in particular,
relates to methods, systems, devices, and computer software for
efficient synthesis of realistic images.
[0004] The invention also relates to the field of systems,
computer-implemented methods and computer program code products for
evaluating integrals, and provides systems, computer-implemented
methods and computer program code products for evaluating integrals
using quasi-Monte Carlo methodologies, and in particular
embodiments, adaptive quasi-Monte Carlo integration and adaptive
integro-approximation in conjunction with techniques including a
scrambled Halton Sequence, stratification by radical inversion,
stratified samples from the Halton Sequence, deterministic
scrambling, bias elimination by randomization, adaptive and
deterministic anti-aliasing, anti-aliasing by rank-1 lattices, and
trajectory splitting by dependent sampling and domain
stratification induced by rank-1 lattices.
BACKGROUND OF THE INVENTION
[0005] The use of synthetic images has become increasingly
important and widespread in motion pictures and other commercial
and scientific applications. A synthetic image represents a
two-dimensional array of digital values, called picture elements or
pixels, and thus can be regarded as a two-dimensional function.
Image synthesis, then, is the process of creating synthetic images
from scenes.
[0006] As a general matter, digital images are generated by
rasterization (as described in greater detail below and in the
references cited in this document, which are incorporated herein by
reference as if set forth in their entireties herein), or, in the
case of photorealistic images of three-dimensional scenes, by ray
tracing (also as described in greater detail below and in the
references cited herein). Both approaches aim at determining the
appropriate color for each pixel by projecting the original
function into the pixel basis. Due to the discrete representation
of the original function, the problem of aliasing arises, as
described below.
[0007] Image synthesis is perhaps the most visible part of computer
graphics. On the one hand it is concerned with physically correct
image synthesis, which intends to identify light paths that connect
light sources and cameras and to sum up their contributions. On the
other hand it also comprises non-photorealistic rendering, such as
the simulation of pen strokes or watercolor.
[0008] The underlying mathematical task of image synthesis is to
determine the intensity I (k, l, t, .lamda.), where (k, l) is the
location of a pixel on the display medium. Computing the intensity
of a single pixel requires an integration function over the pixel
area. This integral is often highly complex, as discussed below,
and cannot be solved analytically, thus requiring numerical methods
for solution, which may include Monte Carlo and quasi-Monte Carlo
methods. In particular, image synthesis is an integro-approximation
problem for which analytical solutions are available only in
exceptional cases. Therefore numerical techniques need to be
applied. While standard graphics textbooks still recommend elements
of classical Monte Carlo integration, the majority of visual
effects in movie industry are produced by using quasi-Monte Carlo
techniques.
[0009] However, typical numerical methods used in such applications
have their own limitations and attendant problems. It would
therefore be desirable to provide improved methods and systems for
image synthesis whereby realistic images can be rendered
efficiently.
[0010] In computer graphics, a computer is used to generate digital
data that represents the projection of surfaces of objects in, for
example, a three-dimensional scene, illuminated by one or more
light sources, onto a two-dimensional image plane, to simulate the
recording of the scene by, for example, a camera. The camera may
include a lens for projecting the image of the scene onto the image
plane, or if may comprise a pinhole camera in which ease no lens is
used. The two-dimensional image is in the form of an array of
picture elements, called "pixels" or "pels," and the digital data
generated for each pixel represents the color and luminance of the
scene as projected onto the image plane at the point of the
respective pixel in the image plane. The surfaces of the objects
may have any of a number of characteristics, including shape,
color, specularity, texture, and so forth, which are preferably
rendered in the image as closely as possible, to provide a
realistic-looking image.
[0011] Generally, the contributions of the light reflected from the
various points in the scene to the pixel value representing the
color and intensity of a particular pixel are expressed in the form
of the one or more integrals of relatively complicated functions.
Since the integrals used in computer graphics generally will not
have a closed-form solution, numerical methods must be used to
evaluate them and thereby generate the pixel value. Typically, a
conventional "Monte Carlo" method has been used in computer
graphics to numerically evaluate the integrals. Generally, in the
Monte Carlo method, to evaluate an integral
<f>=.intg..sub.[0,1).sub.sf(x)dx (1.1)
where f(x) is a real function on the s-dimensional unit cube
[0,1).sup.s, that is, an s-dimensional cube each of whose dimension
includes "zero," and excludes "one." First, a number N of
statistically-independent randomly-positioned points x.sub.i, i=1,
. . . , N, are generated over the integration domain. The random
points x.sub.i are used as sample points for which sample values
f(x.sub.i) are generated for the function f(x), and an estimate f
for the integral is generated as
f .apprxeq. f _ = 1 N i = 1 N f ( x i ) ( 1.2 ) ##EQU00001##
[0012] As the number of random points used in generating the sample
points f(x.sub.i) increases, the value of the estimate f will
converge toward the actual value of the integral <f>.
Generally, the distribution of estimate values that will be
generated for various values of N, that is, for various numbers of
sample points, of being normal distributed around the actual value
with a standard deviation .sigma. which can be estimated by
.sigma. = 1 N - 1 ( f _ 2 - f _ 2 ) ( 1.3 ) ##EQU00002##
if the points x.sub.i used to generate the sample values f(x.sub.i)
are statistically independent, that is, if the points x.sub.i are
truly positioned at random in the integration domain.
[0013] Generally, it has been believed that random methodologies
like the Monte Carlo method are necessary to ensure that
undesirable artifacts, such as Moire patterns and aliasing and the
like, which are not in the scene, will not be generated in the
generated image. However, several problems arise from use of the
Monte Carlo method in computer graphics. First, since the sample
points x.sub.i used in the Monte Carlo method are randomly
distributed, they may clump in various regions over the domain over
which the integral is to be evaluated. Accordingly, depending on
the set of points that are generated, in the Monte Carlo method for
significant portions of the domain there may be no sample points
x.sub.i for which sample values f(x.sub.i) are generated. In that
case, the error can become quite large. In the context of
generating a pixel value in computer graphics, the pixel value that
is actually generated using the Monte Carlo method may not reflect
some elements which might otherwise be reflected if the sample
points x.sub.i were guaranteed to be more evenly distributed over
the domain. This problem can be alleviated somewhat by dividing the
domain into a plurality of sub-domains, hut it is generally
difficult to determine a priori the number of sub-domains into
which the domain should be divided, and, in addition, in a
multi-dimensional integration region, which would actually be used
in computer graphics rendering operations, the partitioning of the
integration domain into sub-domains, which are preferably of equal
size, can be quite complicated.
[0014] In addition, since the method makes use of random numbers,
the error | f-<f>|, where |x| represents the absolute value
of the value x between the estimate value f and actual value
<f> is probabilistic, and, since the error values for various
large values of N are close to normal distribution around the
actual value <f>, only sixty-eight percent of the estimate
values f that might be generated are guaranteed to lie within one
standard deviation of the actual value <f>.
[0015] Furthermore, as is clear from Equation (1.3), the standard
deviation .sigma. decreases with increasing numbers N of sample
points, proportional to the reciprocal of square root of N, that
is,
1 N . ##EQU00003##
Thus, if it is desired to reduce the statistical error by a factor
of two, it will be necessary to increase the number of sample
points N by a factor of four, which, in turn, increases the
computational load that is required to generate the pixel values,
for each of the numerous pixels in the image.
[0016] Additionally, since the Monte Carlo method requires random
numbers to define the coordinates of respective sample points
x.sub.i in the integration domain, an efficient mechanism for
generating random numbers is needed. Generally, digital computers
are provided with so-called "random number generators," which are
computer programs which can be processed to generate a set of
numbers that are approximately random. Since the random number
generators use deterministic techniques, the numbers that are
generated are not truly random. However, the property that
subsequent random numbers from a random number generator are
statistically independent should be maintained by deterministic
implementations of pseudo-random numbers on a computer.
[0017] Grabenstein describes a computer graphics system and method
for generating pixel values for pixels in an image using a strictly
deterministic methodology for generating sample points, which
avoids the above-described problems with the Monte Carlo method.
The strictly deterministic methodology described in Grabenstein
provides a low-discrepancy sample point sequence which ensures, a
priori, that the sample points are generally more evenly
distributed throughout the region over which the respective
integrals are being evaluated. In one embodiment, the sample points
that are used are based on the Halton sequence.
[0018] In a Halton sequence generated for number base b, where base
b is a selected prime number, the k-th value of the sequence,
represented by is generated by use of a "radical inverse" function
.PHI..sub.b that is generally defined as
.PHI. b : N 0 .fwdarw. I i = j = 0 .infin. a j ( i ) b j j = 0
.infin. a j ( i ) b - j - 1 ( 1.4 ) ##EQU00004##
where (a.sub.j).sub.j=0.sup..infin. is the representation of i in
integer base b. Generally, a radical inverse of a value k is
generated by technique including the following steps (1)-(3):
[0019] (1) writing the value k as a numerical representation of the
value in the selected base b, thereby to provide a representation
for the value as D.sub.MD.sub.M-1 . . . D.sub.2 D.sub.1, where
D.sub.m, (m= 1, 2, . . . , M) are the digits of the
representation;
[0020] (2) putting a radix point, corresponding to a decimal point
for numbers written in base ten, at the least significant end of
the representation D.sub.MD.sub.M-1 . . . D.sub.2D.sub.1 written in
step (1) above; and
[0021] (3) reflecting the digits around the radix point to provide
0. D.sub.MD.sub.M-1 . . . D.sub.2 D.sub.1, which corresponds to
H.sub.b.sup.k.
[0022] It will be appreciated that, regardless of the base h
selected for the representation, for any series of values, one,
two, . . . k, written in base b, the least significant digits of
the representation will change at a faster rate than the most
significant digits. As a result, in the Halton sequence
H.sub.b.sup.1, H.sub.b.sup.2, . . . , H.sub.b.sup.k, the most
significant digits will change at the faster rate, so that the
early values in the sequence will be generally widely distributed
over the interval from zero to one, and later values in the
sequence will fill in interstices among the earlier values in the
sequence. Unlike the random or pseudo-random numbers used in the
Monte Carlo method as described above, the values of the Halton
sequence are not statistically independent; on the contrary, the
values of the Halton sequence are strictly deterministic,
"maximally avoiding" each other over the interval, and so they will
not clump, whereas the random or pseudo-random numbers used in the
Monte Carlo method may clump.
[0023] It will be appreciated that the Halton sequence as described
above provides a sequence of values over the interval from zero to
one, inclusive along a single dimension, A multi-dimensional Halton
sequence can be generated in a similar manner, but using a
different base for each dimension, where the bases are relatively
prime.
[0024] A generalized Halton sequence, of which the Halton sequence
described above is a special case, is generated as follows. For
each starting point along the numerical interval from zero to one,
inclusive, a different Halton sequence is generated. Defining the
pseudo-sum x.sym..sub.py for any x and y over the interval from
zero to one, inclusive, for any integer p having a value greater
than two, the pseudo-sum is formed by adding the digits
representing x and y in reverse order, from the most Significant
digit to the least significant digit, and for each addition also
adding in the carry generated from the sum of next more significant
digits. Thus, if x in base b is represented by 0, X.sub.1X.sub.2 .
. . X.sub.M-1X.sub.M, where each X.sub.m is a digit in base b, and
if y in base b is represented by 0, Y.sub.1Y.sub.2 . . .
Y.sub.N-1Y.sub.N, where each Y.sub.n is a digit in base b, where M
is the number of digits in the representation of x in base b, and
where N is the number of digits in the representation of y in base
b, and where M and N may differ, then the pseudo-sum z is
represented by 0, Z.sub.1Z.sub.2 . . . Z.sub.L-1Z.sub.L, where each
Z.sub.1 is a digit hi base b given by
Z.sub.1=(X.sub.1+Y.sub.1+C.sub.1) mod b, where mod represents the
modulo function, and
C l = { 1 for X t - 1 + Y l - 1 + Z l - 1 .gtoreq. b 0 otherwise
##EQU00005##
is a carry value from the 1-1st digit position, with C.sub.1 being
set to zero.
[0025] Using the pseudo-sum function as described above, the
generalized Halton sequence that is used in the system described in
Grabenstein is generated as follows. If b is an integer, and
x.sub.0 is an arbitrary value on the interval from zero to one,
inclusive, then the p-adic von Neumann-Kakutani transformation
T.sub.b(x) is given by
T p ( x ) := x .sym. p 1 b ( 1.5 ) ##EQU00006##
and the generalized Halton sequence x.sub.0, x.sub.1, x.sub.2, . .
. is defined recursively as
x.sub.n+1=T.sub.b(x.sub.n) (1.6)
From Equations (1.5) and (1.6), it is clear that, for any value for
b, the generalized Halton sequence can provide that a different
sequence will be generated for each starting value of x, that is,
for each x.sub.0. It will be appreciated mat the Halton sequence
H.sub.b.sup.k as described above is a special case of the
generalized Halton sequence in Equations (1.5) and (1.6) for
x.sub.0=0.
[0026] The use of a strictly deterministic low-discrepancy sequence
such as the Halton sequence or the generalized Halton sequence can
provide a number of advantages over the random or pseudo-random
numbers that have are used in connection with the Monte Carlo
technique. Unlike the random numbers used in connection with the
Monte Carlo technique, the low discrepancy sequences ensure that
the sample points are more evenly distributed over a respective
region or time interval, thereby reducing error in the image which
can result from clumping of such sample points which can occur in
the Monte Carlo technique. That can facilitate the generation of
images of improved quality when using the same number of sample
points at the same computational cost as in the Monte Carlo
technique.
[0027] It would also be desirable to provide methods and systems
that provide image synthesis by adaptive quasi-Monte Carlo
integration and adaptive integro-approximation in conjunction with
techniques including a scrambled Halton Sequence, stratification by
radical inversion, stratified samples from the Halton Sequence,
deterministic scrambling, bias elimination by randomization,
adaptive and deterministic anti-aliasing, anti-aliasing by rank-1
lattices, and trajectory splitting by dependent sampling and domain
stratification induced by rank-1 lattices.
SUMMARY OF THE INVENTION
[0028] One aspect of the present invention relates to the
generation and synthesis of images, such as for display in a motion
picture or other dynamic display. The invention provides improved
methods and systems for image synthesis, including efficient
methods for determining intensity whereby realistic images can be
rendered efficiently within the limits of available computational
platforms.
[0029] More particularly, the invention provides anew and improved
system and computer-implemented method for evaluating integrals
using quasi-Monte Carlo methodologies, and in particular
embodiments, adaptive quasi-Monte Carlo integration aid adaptive
integro-approximation in conjunction with techniques including a
scrambled Halton Sequence, stratification by radical inversion,
stratified samples from the Halton Sequence, deterministic
scrambling, bias elimination by randomization, adaptive and
deterministic anti-aliasing, anti-aliasing by rank-1 lattices, and
trajectory splitting by dependent sampling and domain
stratification induced by rank-1 lattices.
[0030] In brief summary, the invention provides a computer graphics
system for generating a pixel value for a pixel in an image, the
pixel value being representative of a point in a scene as recorded
on an image plane of a simulated camera, the computer graphics
system comprising a sample point generator and a function
evaluator. The sample point generator is configured to generate a
set of sample points, at least one sample point being generated
using at least one dependent sample, the at least one dependent
sample comprising at least one element of a low-discrepancy
sequence offset by at least one element of another low-discrepancy
sequence. The function evaluator is configured to generate at least
one value representing an evaluation of a selected function at one
of the sample points generated by the sample point generator, the
value generated by the function evaluator corresponding to the
pixel value.
[0031] Another aspect of the invention comprises a computer program
product for use in a computer graphics system, for enabling the
computer graphics system to generate a pixel value for a pixel in
an image displayable via a display device, the pixel value being
representative of a point in a scene, the computer program product,
comprising a computer-readable medium having encoded thereon:
[0032] A. computer-readable program instructions executable to
enable the computer graphics system to generate a set of sample
points, at least one sample point being generated using at least
one sample, the at least one sample comprising at least one element
of a low-discrepancy sequence, and wherein the set of sample points
comprises quasi-Monte Carlo points of low discrepancy; and
[0033] B. computer-readable program instructions executable to
enable the computer graphics system to evaluate a selected function
at one of the sample points to generate a value, the generated
value corresponding to the pixel value, the pixel value being
usable to generate a display-controlling electronic output.
BRIEF DESCRIPTION OF THE DRAWINGS
[0034] This invention is pointed out with particularity in the
appended claims. The above and further advantages of this invention
may be better understood by referring to the following description
taken in conjunction with the accompanying drawings, in which:
[0035] FIG. 1 shows a diagram of a computer graphics system
suitable for use in implementations of various aspects of the
invention described herein.
[0036] FIG. 2 shows a diagram illustrating components of the
computer graphics system and processor module shown in FIG. 1.
[0037] FIG. 2A shows a diagram illustrating further components of a
computing system according to aspects of the present invention.
[0038] FIG. 3 shows a diagram of a network configuration suitable
for use in implementations of various aspects of the invention
described herein.
[0039] FIG. 4 shows a diagram illustrating components of the of the
network configuration shown in FIG. 3.
[0040] FIGS. 5 and 6 show code fragments for generating imaging
data in accordance with aspects of the invention.
[0041] FIGS. 7A and 7B show plots of the first two components of
the Halton sequence.
[0042] FIGS. 8A and 8B show a pair of low-dimensional projections,
including the Halton sequence and the scrambled Halton sequence for
designated points.
[0043] FIG. 9 shows a plot of a sample pattern that is tiled over
the image plane.
[0044] FIG. 10 shows a plot illustrating an interleaved adaptive
supersampling technique according to a further aspect of the
invention.
[0045] FIGS. 11A-H show a series of plots illustrating classical
quasi-Monte Carlo points, along with their mutual minimum
distance.
[0046] FIGS. 12A-C shows a series of drawings illustrating
selection of lattices by maximum minimum distance.
[0047] FIG. 13 shows a computer-generated image of an infinite
plane with a checkerboard texture.
[0048] FIGS. 14A-C are a series of alternative sampling patterns of
computer graphics.
[0049] FIGS. 15A-E show sampling patterns using quasi-Monte Carlo
points.
[0050] FIG. 16 shows an illustration of how samples in a pixel are
determined by tiled instances of a Hammersley point set.
[0051] FIGS. 17A and 17B are plots illustrating how samples from
the Halton sequence in the unit square are scaled to fit the pixel
raster.
[0052] FIGS. 18A-C are a series of plots illustrating replications
by rank-1 lattices.
[0053] FIGS. 19-22 show a series of flowcharts of general methods
according to aspects of the invention.
DETAILED DESCRIPTION OF THE INVENTION
[0054] The present invention relates to the generation and
synthesis of images, such as for display in a motion picture or
other dynamic display. The techniques described herein are
practiced as part of a computer graphics system, in which a pixel
value is generated for each pixel in an image. The pixel value is
representative of a point in a scene as recorded on an image plane
of a simulated camera. The computer graphics system is configured
to generate the pixel value for an image using a selected
methodology.
[0055] The following discussion describes methods, structures,
systems, and display technology in accordance with these
techniques. It will be understood by those skilled in the art that
the described methods and systems can be implemented in software,
hardware, or a combination of software and hardware, using
conventional computer apparatus such as a personal computer (PC) or
equivalent device operating in accordance with (or emulating) a
conventional operating system such as Microsoft Windows, Linux, or
Unix, either in a standalone configuration or across a network. The
various processing means and computational means described below
and recited in the claims may therefore be implemented in the
software and/or hardware elements of a properly configured digital
processing device or network of devices. Processing may be
performed sequentially or in parallel, and may be implemented using
special purpose or reconfigurable hardware.
[0056] FIG. 1 attached hereto depicts an illustrative computer
system 10 that makes use of such a strictly deterministic
methodology. With reference to FIG. 1, the computer system 10 in
one embodiment includes a processor module 11 and operator
interface elements comprising operator input components such as a
keyboard 12A and/or a mouse 12B (generally identified as operator
input element(s) 12) and an operator output element such as a video
display device 13. The illustrative computer system 10 is of the
conventional stored-program computer architecture. The processor
module 11 includes, for example, one or more processor, memory and
mass storage devices, such as disk and/or tape storage elements
(not separately shown), which perform processing and storage
operations in connection with digital data provided thereto. The
operator input element(s) 12 are provided to permit an operator to
input information for processing. The video display device 13 is
provided to display output information generated by the processor
module 11 on a screen 14 to the operator, including data that the
operator may input for processing, information that the operator
may input to control processing, as well as information generated
during processing. The processor module 11 generates information
for display by the video display device 13 using, a so-called
"graphical user interface" ("GUI"), in which information for
various applications programs is displayed using various "windows."
Although the computer system 10 is shown as comprising particular
components, such as the keyboard 12A and mouse 12B for receiving
input information from an operator, and a video display device 13
for displaying output information to the operator, it will be
appreciated that the computer system 10 may include a variety of
components in addition to or instead of those depicted in FIG.
1.
[0057] In addition, the processor module 11 includes one or more
network ports, generally identified by reference numeral 14, which
are connected to communication links which connect the computer
system 10 in a computer network. The network ports enable the
computer system 10 to transmit information to, and receive
information from, other computer systems and other devices in the
network. In a typical network organized according to, for example,
the client-server paradigm, certain computer systems in the network
are designated as servers, which store data and programs
(generally, "information") for processing by the other, client
computer systems, thereby to enable the client computer systems to
conveniently share the information. A client computer system which
needs access to information maintained by a particular server will
enable the server to download the information to it over the
network. After processing the data, the client computer system may
also return the processed data to the server for storage. In
addition to computer systems (including the above-described servers
and clients), a network may also include, for example, printers and
facsimile devices, digital audio or video storage and distribution
devices, and the like, which may be shared among the various
computer systems connected in the network. The communication links
interconnecting the computer systems in the network may, as is
conventional, comprise any convenient information-earning medium,
including wires, optical fibers or other media for carrying signals
among the computer systems. Computer systems transfer information
over the network by means of messages transferred over the
communication links, with each message including information and an
identifier identifying the device to receive the message.
[0058] FIG. 2 shows a diagram illustrating the sample point
generator 20, function evaluator 22, simulated rays 24 and
simulated lens 26 processing aspects of a computer graphics system
10 and processor module 11 in accordance with the invention.
[0059] FIG. 2A shows a diagram illustrating additional components
of the computing system 10 according, to further aspects of the
invention, described below. As shown in FIG. 2A, the computing
system 10 further includes a sample point generator 30 for
generating a set of sample points, at least one sample point being
generated using at least one sample, the at least one sample
comprising at least one element of a sequence, and wherein the set
of sample points comprises quasi-Monte Carlo points. The computing
system 10 also includes a function evaluator 32 in communication
with the sample point generator 30 for evaluating a selected
function at one of the sample points to generate a value, the
generated value corresponding to the pixel value. The pixel value
is usable to generate an electronic output, that controls the
display 13.
[0060] In addition to the computer system 10 shown in FIGS. 1 and
2, methods, devices or software products in accordance with the
invention can operate on any of a wide range of conventional
computing devices and systems, such as those depicted by way of
example in FIG. 3 (e.g. network system 100), whether standalone,
networked, portable or fixed, including conventional PCs 102,
laptops 104, handheld or mobile computers 106, or across the
Internet or other networks 108, which may in turn include servers
110 and storage 112.
[0061] In line with conventional computer software and hardware
practice, a software application configured in accordance with the
invention can operate within, e.g., a PC 102 like that shown in
FIG. 4, in which program instructions can be read from CD ROM 116,
magnetic disk or other storage 120 and loaded into RAM 114 for
execution by CPU 118. Data can be input into the system via any
known device or means, including a conventional keyboard, scanner,
mouse or other elements 103.
[0062] Those skilled in the art will understand that the method
aspects of the invention described below can be executed in
hardware elements, such as an Application-Specific Integrated
Circuit (ASIC) constructed specifically to carry out the processes
described herein, using ASIC construction techniques known to ASIC
manufacturers. Various forms of ASICs are available from many
manufacturers, although currently available ASICs do not provide
the functions described in this patent application. Such
manufacturers include Intel Corporation and NVIDIA Corporation,
both of Santa Clara, Calif. The actual semiconductor elements of
such ASICs and equivalent integrated circuits are not part of the
present invention, and will not be discussed in detail herein.
[0063] Those skilled in the art will also understand that method
aspects of the present invention can be carried out within
commercially available digital processing systems, such as
workstations and personal computers (PCs), operating under the
collective command of the workstation or PC's operating system and
a computer program product configured in accordance with the
present invention. The term "computer program product" can
encompass any set of computer-readable programs instructions
encoded on a computer readable medium. A computer readable medium
can encompass any form of computer readable element, including, but
not limited to, a computer hard disk, computer floppy disk,
computer-readable flash drive, computer-readable RAM or ROM element
or any other known means of encoding, storing or providing digital
information, whether local to or remote from the workstation, PC or
other digital processing device or system. Various forms of
computer readable elements and media are well known in the
computing arts, and their selection is left to the implementer. In
each case, the invention is operable to enable a computer system to
calculate a pixel value, and the pixel value can be used by
hardware elements in the computer system, which can be conventional
elements such as graphics cards or display controllers, to generate
a display-controlling electronic output. Conventional graphics
cards and display controllers are well known in the computing arts,
are not necessarily part of the present invention, and their
selection can be left to the implementer.
[0064] In particular, the systems illustrated in FIGS. 1-4 may be
used, in accordance with the following described aspects of the
invention, to implement a computer graphics system that evaluates
integrals using a quasi-Monte Carlo methodology, which can include
adaptive quasi-Monte Carlo integration and adaptive
integro-approximation in conjunction with techniques including a
scrambled Halton Sequence, stratification by radical inversion,
stratified samples from the Halton Sequence, deterministic
scrambling, bias elimination by randomization, adaptive and
deterministic anti-aliasing; anti-aliasing by rank-1 lattices, and
trajectory splitting by dependent sampling and domain
stratification induced by rank-1 lattices.
[0065] Various aspects, examples, features, embodiments and
practices in accordance with the present invention will be set
forth hi detail in the present Detailed Description of the
invention, which is organized into the following sections:
I. Introduction, Overview and Description of Quasi-Monte Carlo
Methodologies in Which Sample Points Represent Dependent Samples
Generated Using a Low-Discrepancy Sequence
II. Image Synthesis by Adaptive Quasi-Monte Carlo Integration
III. Additional Examples and Points Regarding Quasi-Monte Carlo
Integration
IV. General Methods
[0066] The present application is a continuation-in-part of
pending, commonly owned U.S. patent application Ser. No. 10/299,958
filed Nov. 19, 2002 (Attorney Docket MENT-072, inventor: Alexander
Keller), entitled "System and Computer-Implemented Method for
Evaluating Integrals Using a Quasi-Monte Carlo Methodology in Which
Sample Points Represent Dependent Samples Generated Using a
Low-Discrepancy Sequence," and the detailed description of the
present invention begins by setting forth salient points from that
application.
I. Introduction, Overview and Description of Quasi-Monte Carlo
Methodologies in which Sample Points Represent Dependent Samples
Generated Using a Low-Discrepancy Sequence
[0067] Aspects of the present invention provide a computer graphic
system and method for generating pixel values for pixels in an
image of a scene, which makes use of a strictly deterministic
quasi-Monte Carlo methodology in conjunction with various
sub-techniques, which can include, for example, trajectory
splitting by dependent sampling for generating sample points for
use in generating sample values for evaluating the integral or
integrals whose function(s) represent the contributions of the
light reflected from the various points in the scene to the
respective pixel value, rather than the random or pseudo-random
Monte Carlo methodology which has been used in the past. The
strictly deterministic methodology ensures a priori that the sample
points will be generally more evenly distributed over the interval
or region over which the integral(s) is (are) to be evaluated in a
low-discrepancy manner.
[0068] It will be helpful to initially provide some background on
operations performed by the computer graphics system in generating
an image. Generally, the computer graphic system generates an image
that attempts to simulate an image of a scene that would be
generated by a camera. The camera includes a shutter that will be
open for a predetermined time T starling at a time to allow light
from the scene to be directed to an image plane. The camera may
also include a lens or lens model (generally, "lens") that serves
to focus light from the scene onto the image plane. The average
radiance flux L.sub.m,n through a pixel at position (m, n) on an
image plane P, which represents the plane of the camera's recording
medium, is determined by
L m , n = 1 A P T A L .intg. A P .intg. t 0 t 0 + T .intg. A L L (
h ( x , t , y ) , - .omega. ( x , t , y ) ) f m , n ( x , y , t ) y
t x ( 1.7 ) ##EQU00007##
where A.sub.p refers to the area of the pixel, A.sub.L refers to
the area of the portion of the lens through which rays of light
pass from the scene to the pixel, and f.sub.m,n represents a
filtering kernel associated with the pixel. An examination of the
integral in Equation (1.7) will reveal that, for the variables of
integration, x, y and t, the variable y refers to integration over
the lens area A.sub.L, the variable t refers to integration over
time (the time interval from t.sub.0 to t.sub.0+T) and the variable
x refers to integration over the pixel area A.sub.p.
[0069] The value of the integral in Equation (1.7) is approximated
in accordance with a quasi-Monte Carlo methodology by identifying
N.sub.p sample points x.sub.i in the pixel area, and, for each
sample point, shooting N.sub.T rays at limes in the time interval
to t.sub.0 to t.sub.0+T through the focus into the scene, with each
ray spanning N.sub.L sample points y.sub.i,j,k on the lens area
A.sub.L. The manner in which subpixel jitter positions x.sub.i,
points in time t.sub.i,j and positions on the lens y.sub.i,j,k are
determined will be described below. These three parameters
determine the primary ray hitting the scene geometry in h(x.sub.i,
t.sub.i,j, y.sub.i,j,k) with the ray direction .omega.(x.sub.i,
ti,j, yi,j,k). In this manner, the value of the integral in
Equation (1.7) can be approximated as follows:
L m , n .apprxeq. 1 N i = 0 N p - 1 1 N T j = 0 N T - 1 1 N L k = 0
N L - 1 L ( h ( x i , t i , j , y i , j , k ) , - .omega. ( x i , t
i , j y i , j , k ) ) f m , n ( x i , t i , j , y i , j , k ) , (
1.8 ) ##EQU00008##
where N is the total number of rays directed at the pixel.
[0070] It will be appreciated that rays directed from the scene
toward the image plane can comprise rays directly from one or more
light sources in the scene, as well as rays reflected off surfaces
of objects in the scene. In addition, it will be appreciated that a
ray that is reflected off a surface may have been directed to the
surface directly from a light source, or a ray that was reflected
off another surface. For a surface that reflects light rays, a
reflection operator T.sub.fp is defined that includes a diffuse
portion T.sub.fd, a glossy portion T.sub.fg and a specular portion
T.sub.fs, or
T.sub.f.sub.g=T.sub.f.sub.d+T.sub.f.sub.g+T.sub.f.sub.s (1.9)
[0071] In that case, the Fredholm integral equation
L=L.sub.e+T.sub.frL governing light transport can be represented
as
L=L.sub.e+T.sub.f.sub.r.sub.-f.sub.sL.sub.e+T.sub.f.sub.s(L-L.sub.e)+T.s-
ub.f.sub.sL+T.sub.f.sub.g.sub.f+.sub.sL+T.sub.f.sub.dT.sub.f.sub.dL
(1.10)
where transparency has beet) ignored for the sake of simplicity;
transparency is treated in an analogous manner. The individual
terms in Equation (1.10) are as follows; [0072] (i) L.sub.e
represents flux due to a light source; [0073] (ii)
T.sub.f.sub.g.sub.-f.sub.gL.sub.e (where
T.sub.f.sub.g.sub.-f.sub.g-T.sub.f.sub.g=-T.sub.f.sub.g) represents
direct illumination, that is, flux reflected off a surface that was
provided thereto directly by a light source; the specular
component, associated with the specular portion T.sub.fs of the
reflection operator, will be treated separately since it is modeled
using a .delta.-distribution: [0074] (iii) T.sub.f.sub.g
(L-L.sub.g) represents glossy illumination, which is handled by
recursive distribution ray tracing, where, in the recursion, the
source illumination has already been accounted for by the direct
illumination (item (ii) above): [0075] (iv) T.sub.f.sub.gL
represents a specular component, which is handled by recursively
using L for the reflected ray; [0076] (v)
T.sub.f.sub.gT.sub.f.sub.g.sub.+f.sub.gL (where
T.sub.f.sub.g.sub.+f.sub.g=T.sub.f.sub.g+T.sub.f.sub.g) represents
a caustic component, which is a ray that has been, reflected off a
glossy or specular surface (reference the
T.sub.f.sub.g.sub.+f.sub.g operator) before bitting a diffuse
surface (reference the T.sub.f.sub.g operator); this contribution
can be approximated by a high resolution caustic photon map; and
[0077] (vi) T.sub.f.sub.gT.sub.f.sub.dL represents ambient light,
which is very smooth and is therefore approximated using a low
resolution global photon map.
[0078] As noted above, the value of the integral in Equation (1.7)
is approximated by solving Equation (1.8) making use of sample
points x.sub.i, t.sub.i,j, and y.sub.i,j,k, where x.sub.i refers to
sample points within area A.sub.L of the respective pixel at
location (m, n) in the image plane, t.sub.i,j refers to sample
points within the time interval t.sub.0 to t.sub.0+T during which
the shutter is open, and y.sub.i,j,k refers to sample points on the
lens A.sub.L. In accordance with one aspect of the invention, the
sample points x.sub.i comprise two-dimensional Hammersley points,
which are defined as
( i N , .PHI. 2 ( i ) ) ##EQU00009##
where 0.ltoreq.i<N, and .PHI..sub.2(i) refers to the radical
inverse of i in base two. Generally, the s-dimensional Hammersley
point set is defined as follows:
U N , s Hammersley : [ 0 , , N - 1 ] I s i x i := ( i N , .PHI. b 1
( i ) , , .PHI. b s - 1 ( i ) ) ( 1.11 ) ##EQU00010##
where 1.sup.s is the 5-dimensional unit cube [0,1).sup.s (that is,
an s-dimensional cube each of whose dimensions includes zero, and
excludes one), the number of points N in the set is fixed and
b.sub.1, . . . , b.sub.g-1 are bases. The bases do not need to be
prime numbers, but they are preferably relatively prime to provide
a uniform distribution. The radical inverse function .PHI..sub.b,
in turn, is generally defined as
.PHI. b : N 0 I i = j = 0 .infin. a j ( i ) b j j = 0 .infin. a j (
i ) b - j - 1 ( 1.12 ) ##EQU00011##
where (a.sub.j).sub.j=0.sup..infin. is the representation of i in
integer base b. At N=(2.sup.n).sup.2, the two-dimensional
Hammersley points are a (0, 2n, 2)-net in base two, which are
stratified on a 2.sup.n.times.2.sup.n grid and a Latin hypercube
sample at the same time. Considering the grid as subpixels, the
complete subpixel grid underlying the image plane can be filled by
simply abutting copies of the grid to each other.
[0079] Given integer subpixel coordinates (s.sub.x, s.sub.y) the
instance i and coordinates (x,y) for the sample point x.sub.i in
the image plane can be determined as follows.
Preliminary, examining
( i N , .PHI. 2 ( i ) ) i = 0 N - 1 ##EQU00012##
one observes the following: [0080] (a) each line in the stratified
pattern is a shifted copy of another, and [0081] (b) the pattern is
symmetric to the line y=x, that is, each column is a shifted copy
of another column.
[0082] Accordingly, given the integer permutation
.sigma.(k):=2.sup.n.PHI..sub.2(k) for 0.ltoreq.k<2.sup.n,
subpixel coordinates (s.sub.x, s.sub.y) are mapped onto strata
coordinates (j, k):=(s.sub.x mod 2.sup.n, s.sub.y mod 2.sup.n), an
instance number i is computed as
i=j2.sup.n+.sigma.(k) (1.13)
and fee positions of the jittered subpixel sample points are
determined according to
x j = ( s x + .PHI. 2 ( k ) , s y + .PHI. 2 ( j ) ) = ( s x +
.sigma. ( k ) 2 n , s y + .sigma. ( j ) 2 n ) ( 1.14 )
##EQU00013##
[0083] An efficient algorithm for generating the positions of the
jittered subpixel sample points x.sub.i will be provided below in
connection with Code Segment 1. A pattern of sample points whose
positions are determined as described above in connection with
Equations (1.13) and (1.14) has an advantage of having much reduced
discrepancy ever a pattern determined using a Halton sequence or
windowed Halton sequence, as described in Grabenstein, and
therefore the approximation described above in connection with
Equation (1.8) gives in general a better estimation to the value of
the integral described above in connection with Equation (1.7). In
addition, if N is sufficiently large, sample points in adjacent
pixels will have different patterns, reducing the likelihood that
undesirable artifacts will be generated in the image.
[0084] A "ray tree" is a collection of paths of light rays that are
traced from a point on the simulated camera's image plane into fee
scene. The computer graphics system 10 generates a ray tree by
recursively following transmission, subsequent reflection and
shadow rays using trajectory splitting. In accordance with another
aspect of the invention, a path is determined by the components of
one vector of a global generalized scrambled Hammersley point set.
Generally, a scrambled Hammersley point set reduces or eliminates a
problem that can arise in connection with higher-dimensioned
low-discrepancy sequences since the radical inverse function
.PHI..sub.b typically has subsequences of b-1 equidistant values
spaced by 1/b. Although these correlation patterns are merely
noticeable in the full s-dimensional space, they are undesirable
since they are prone to aliasing. The computer graphics system 10
attenuates this effect by scrambling, which corresponds to
application of a permutation to the digits of the b-ary
representation used in the radical inversion. For the symmetric
permutation .sigma. from the symmetric group S.sub.b over integers
0, . . . , b-1, the scrambled radical inverse is defined as
.PHI. b : N 0 .times. S b I ( i , .sigma. ) j = 0 .infin. .sigma. (
a j ( i ) ) b - j - 1 i = j = 0 .infin. a j ( i ) b j ( 1.15 )
##EQU00014##
[0085] If the symmetric permutation .sigma. is the identity, the
scrambled radical inverse corresponds to the unscrambled radical
inverse, in one embodiment the computer graphics system 10
generates the symmetric permutation a recursively as follows.
Starting from the permutation .sigma..sub.2=(0, 1) for base b=2,
the sequence of permutations is defined as follows:
[0086] (i) if the base b is even, the permutation ah is generated
by first taking the values of 2.sigma..sub.b/2 and appending the
values of
2 .sigma. b 2 + 1 ##EQU00015##
and
[0087] (ii) if the base b is odd, the permutation .sigma..sub.b is
generated by taking the values of .sigma..sub.b-1, incrementing
each value that is greater than or equal to
b - 1 2 ##EQU00016##
by one, and inserting the value
b - 1 2 ##EQU00017##
in the middle.
[0088] This recursive procedure results in
.sigma..sub.2=(0,1)
.sigma..sub.3=(0,1,2)
.sigma..sub.4=(0,2,1,3)
.sigma..sub.5=(0,3,2,1,4)
.sigma..sub.6=(0,2,4,1,3,5)
.sigma..sub.7=(0,4,2,6,1,5,3,7)
.sigma..sub.8=(0,4,2,6,1,5,3,7) . . .
[0089] The computer graphics system 10 can generate a generalized
low-discrepancy point set as follows. It is often possible to
obtain a low-discrepancy sequence by taking any rational
s-dimensional point x as a starting point and determine a successor
by applying the corresponding incremental radical inverse function
to the components of x. The result is referred to as the
generalized low-discrepancy point set. This can be applied to both
the Halton sequence and the Hammersley sequence. In the case of the
generalized Halton sequence, this can be formalized as
x.sub.i(.PHI..sub.b.sub.1(i+i.sub.1),.PHI..sub.b.sub.2(i+i.sub.2),
. . . , .PHI.(i+i.sub.s)) (1.16)
where the integer vector (i.sub.1, i.sub.2, . . . , i.sub.s)
represents the offsets per component and is fixed in advance for a
generalized sequence. The integer vector can be determined by
applying the inverse of the radical inversion to the starting point
x. A generalized Hammersley sequence can be generated in an
analogous manner.
[0090] Returning to trajectory splitting, generally trajectory
splitting is the evaluation of a local integral, which is of small
dimension and which makes the actual integrand smoother, which
improves overall convergence. Applying replication, positions of
low-discrepancy sample points are determined that can be used in
evaluating the local integral. The low-discrepancy sample points
are shifted by the corresponding elements of the global scrambled
Hammersley point set. Since trajectory splitting can occur multiple
times on the same level in the ray tree, branches of the ray tree
are decorrelated in order to avoid artifacts, the decorrelation
being accomplished by generalizing the global scrambled Hammersley
point set.
[0091] An efficient algorithm for generating a ray tree will be
provided below in connection with Code Segment 2. Generally, in
that algorithm, the instance number i of the low-discrepancy
vector, as determined above in connection with Equation (1.13), and
the number d of used components, which corresponds to the current
integral dimension, are added to the data structure that is
maintained for the respective ray in the ray tree. The ray tree of
a subpixel sample is completely specified by the instance number i.
After the dimension has been set to "two," which determines the
component of the global Hammersley point set that is to be used
next, the primary ray is cast into the scene to span its ray tree.
In determining the deterministic splitting by the components of low
discrepancy sample points, the computer graphics system 10
initially allocates the required number of dimensions .DELTA.d. For
example, in simulating glossy scattering, the required number of
dimensions will correspond to "two." Thereafter, the computer
graphics system 10 generates scattering directions from the offset
given by the scrambled radical inverses
.PHI..sub.b.sub.d(i,.sigma..sub.b.sub.d), . . . ,
.PHI..sub.b.sub.d+.DELTA.d-1(i,.sigma..sub.b.sub.d+.DELTA.d-1)
yielding the instances
( y i , j ) j = 0 M - 1 = ( .PHI. b d ( i , .sigma. b d ) .sym. J M
, , .PHI. b d + .DELTA. d - 1 ( i , .sigma. b d + .DELTA. d - 1 )
.sym. .PHI. b d + .DELTA. d - 2 ( j , .sigma. b d + .DELTA. d - 2 )
) ( 1.17 ) ##EQU00018##
where ".sym." refers to "addition modulo one." Each direction of
the M replicated rays is determined by y.sub.i,j and enters the
next level of the ray tree with d':=d+.DELTA.d as the new integral
dimension in order to use the next elements of the low-discrepancy
vector, and i'=i+j in order to decorrelate subsequent trajectories.
Using an infinite sequence of low-discrepancy sample points, the
replication heuristic is turned into an adaptive consistent
sampling arrangement. That is, computer graphics system 10 can fix
the sampling rate .DELTA.M, compare current and previous estimates
even .DELTA.M samples, and, if the estimates differ by less than a
predetermined threshold value, terminate sampling. The computer
graphics system 10 can, in turn, determine the threshold value, by
importance information, that is, how much the local integral,
contributes to the global integral.
[0092] As noted above, the integral described above in connection
with Equation (1.7) is over a finite time period T from t.sub.0 to
t.sub.0+T, during which time the shutter of the simulated camera is
open. During the time period, if an object in the scene moves, the
moving object may preferably be depicted in the image as blurred,
with the extent of blurring being a function of the object's motion
and the time interval t.sub.0+T. Generally, motion during the time
an image is recorded is linearly approximated by motion vectors, in
which case the integrand in Equation (1.7) is relatively smooth
over the time the shutter is open and is suited for correlated
sampling. For a ray instance i, started at the subpixel position
x.sub.i, the offset .PHI..sub.3(i) into the time interval is
generated and the N.sub.T-1 subsequent samples
.PHI. 3 ( i ) + j N T ##EQU00019##
mod 1 are generated for 0<j<N.sub.T, that is
t i , j := t 0 ( .PHI. 3 ( i ) .sym. j N T ) T ( 1.18 )
##EQU00020##
[0093] It will be appreciated that the value of N.sub.T may be
chosen to be "one," in which case there will be no subsequent
samples for ray instance i. Determining sample points in this
mariner fills the sampling space, resulting in a more rapid
convergence to the value of the integral in Equation (1.7). For
subsequent trajectory splitting, rays are decorrelated by setting
the instance i'=i+j.
[0094] In addition to determining the position of the jittered
subpixel sample point x.sub.i, and adjusting the camera and scene
according to the sample point t.sub.i,j for the time, the computer
graphics system also simulates depth of field. In simulating depth
of field, the camera to be simulated is assumed to be provided with
a lens having known optical characteristics and, using geometrical
optics, the subpixel sample point x.sub.i is mapped through the
lens to yield a mapped point x.sub.i'. The lens is sampled by
mapping the dependent samples
y i , j , k = ( ( .PHI. 5 ( i + j , .sigma. 5 ) .sym. k N L ) , (
.PHI. 7 ( i + j , .sigma. 7 ) .sym. .PHI. 2 ( k ) ) ) ( 1.19 )
##EQU00021##
onto the lens area A.sub.L using a suitable one of a plurality of
known transformations. As with N.sub.T, the value of N.sub.L may be
chosen to be "one." Thereafter, a ray is shot from the sample point
on the lens specified by y.sub.i,j,k through the point x.sub.i'
into the scene. The offset
(.PHI..sub.5(i+j,.sigma..sub.5),.PHI..sub.7(i+j,.sigma..sub.7))
in Equation (1.19) comprises the next components taken from fee
generalized scrambled Hammersley point set, which, for trajectory
splitting, is displaced by the elements
( k N L , .PHI. 2 ( k ) ) ##EQU00022##
of the two-dimensional Hammersley point set. The instance of the
ray originating from sample point y.sub.i,j,k is set to i+j+k in
order to decorrelate further splitting down the ray tree. In
Equation (1.19), the scrambled samples (.PHI..sub.5(i+j,
.sigma..sub.5), .PHI..sub.7(i+j, .sigma..sub.7)) are used instead
of the unscrambled samples of (.PHI..sub.5(i+j), .PHI..sub.7(i+j)
since in bases five and seven, up to five unscrambled samples will
lie on a straight line, which will not be the case for fee
scrambled samples.
[0095] In connection with determination of a value for the direct
illumination (T.sub.f.sub.g.sub.-f.sub.gL.sub.e above), direct
illumination is represented as an integral over the surface of the
scene .differential.V, which integral, is decomposed into a sum of
integrals, each over one of the L single area light sources in the
scene. The individual integrals in turn are evaluated by dependent
sampling, that is
( T f r - fs L e ) ( y , z ) = .intg. .differential. V L e ( x , y
) f r ( x , y , z ) G ( x , y ) x = k = 1 L .intg. supp L e , k L e
( x , y ) f r ( x , y , z ) G ( x , y ) x .apprxeq. k = 1 L 1 M K j
= 0 M k - 1 L e ( x j , y ) f r ( x j , y , z ) G ( x j , y ) (
1.20 ) ##EQU00023##
where suppL.sub.e,k refers to the surface of the respective k-th
light source, in evaluating the estimate of the integral for the
k-th light source, for the M.sub.k-th query ray, shadow rays
determine the fraction of visibility of the area light source,
since the point visibility vanes much more than the smooth shadow
effect. For each light source, the emission L.sub.e is attenuated
by a geometric term G, which includes the visibility, and the
surface properties are given by a bidirectional distribution
function f.sub.r-f.sub.s. These integrals are local integrals in
the ray tree yielding the value of one node in the ray tree, and
can be efficiently evaluated using dependent sampling. In dependent
sampling, the query ray comes with the next free integral dimension
d and the instance i from which the dependent samples are
determined in accordance with
x j = ( .PHI. b d ( i , .sigma. b d ) .sym. j M k , .PHI. b d + 1 (
i , .sigma. b d + 1 ) .sym. .PHI. 2 ( j ) ) ( 1.21 )
##EQU00024##
The offset
(.PHI..sub.b.sub.d(i,.sigma..sub.b.sub.d),.PHI..sub.b.sub.d+1(i,.sigma..-
sub.b.sub.d+1))
again is taken from the corresponding generalized scrambled
Hammersley point set, which shifts the two-dimensional Hammersley
point set
( j M k , .PHI. 2 ( j ) ) ##EQU00025##
on the light source. Selecting the sample rate Mk=2.sup.n.sup.y as
a power of two, the local minima are obtained for the discrepancy
of the Hammersley point set that perfectly stratifies the light
source. As an alternative, the light source can be sampled using an
arbitrarily-chosen number M.sub.k of sample points using the
following replication rule:
( j M k , .PHI. M k ( j , .sigma. M k ) ) j = 0 M k - 1
##EQU00026##
Due to the implicit stratification of the positions of the sample
points as described above, the local convergence will be very
smooth.
[0096] The glossy contribution T.sub.f.sub.g(L-L.sub.g) is
determined in a manner similar to that described above in
connection with area light sources (Equations (1.20) and (1.21)),
except that a model f.sub.g used to simulate glossy scattering will
be used instead of the bidirectional distribution function f.sub.r.
In determining the glossy contribution, two-dimensional Hammersley
points are generated for a fixed splitting rate M and shifted
modulo "one" by the offset
(.PHI..sub.b.sub.d(i,.sigma..sub.b.sub.d),.PHI..sub.b.sub.dg(i,.sigma..s-
ub.b.sub.dg))
taken from the current ray tree depth given by the dimension field
d of the incoming ray. The ray trees spanned into the scattering
direction are decorrelated by assigning the instance fields i'=i+j
in a manner similar to that done for simulation of motion blur and
depth of field, as described above. The estimates generated for all
rays are averaged by the splitting rate M and propagated up the ray
tree.
[0097] Volumetric effects are typically provided by performing a
line integration along respective rays from their origins to the
nearest surface point in the scene, in providing for a volumetric
effect the computer graphics system 10 generates from the ray data
a corresponding offset .PHI..sub.b.sub.g(i) which it then uses to
shift the M equidistant samples on the unit interval seen as a
one-dimensional torus. In doing so, the rendering time is reduced
in comparison to use of an uncorrelated jittering methodology. In
addition, such equidistant shifted points typically obtain the best
possible discrepancy in one dimension.
[0098] Global illumination includes a class of optical effects,
such as indirect illumination, diffuse and glossy
inter-reflections, caustics and color bleeding, that the computer
graphics system 10 simulates in generating an image of objects in a
scene. Simulation of global illumination typically involves the
evaluation of a rendering equation. For the general form of an
illustrative rendering equation useful in global illumination
simulation, namely:
L({right arrow over (x)},{right arrow over (w)})=L.sub.e({right
arrow over (x)},{right arrow over (w)})+.intg..sub.S'f({right arrow
over (x)},{right arrow over (w)}.fwdarw.{right arrow over
(w)})G({right arrow over (x)},{right arrow over (x)}')V({right
arrow over (x)},{right arrow over (x)}')L({right arrow over
(x)},{right arrow over (w)}')dA' (1.22)
it is recognized that the light radiated at a particular point
{right arrow over (x)} in a scene is generally the sum of two
components, namely, the amount of light, if any, that is emitted
from the point and the amount of light, if any, that originates
from all other points and which is reflected or otherwise scattered
from the point {right arrow over (x)}. In Equation (1.22), L
({right arrow over (x)}, {right arrow over (w)}) represents the
radiance at the point {right arrow over (x)} in the direction
{right arrow over (w)}=(.theta., .phi.) (where .theta. represents
the angle of direction {right arrow over (w)} relative to a
direction orthogonal, of the surface of the object in the scene
containing the point {right arrow over (x)}, and .phi. represents
the angle of the component of direction {right arrow over (w)} in a
plane tangential to the point {right arrow over (x)}). Similarly, L
({right arrow over (x)}', {right arrow over (w)}') in the integral
represents the radiance at the point {right arrow over (x)}' in the
direction {right arrow over (w)}' (.theta.', .phi.') (where
.theta.' represents the angle of direction {right arrow over (w)}'
relative to a direction orthogonal of the surface of the object in
the scene containing the point {right arrow over (x)}', and .phi.'
represents the angle of the component of direction {right arrow
over (w)}' in a plane tangential to the point {right arrow over
(x)}'), and represents the light, if any, that is emitted from
point {right arrow over {right arrow over (x)}' which may be
reflected or otherwise scattered from point {right arrow over
(x)}.
[0099] Continuing with Equation (1.22), L.sub.e ({right arrow over
(x)}, {right arrow over (w)}) represents the first component of the
sum, namely, the radiance due to emission from the point {right
arrow over (x)} in the direction {right arrow over (w)}, and the
integral over the sphere S' represents the second component,
namely, the radiance due to scattering of light at point {right
arrow over (x)}. f({right arrow over (x)}, {right arrow over
(w)}'.fwdarw.{right arrow over (w)}) is a bidirectional scattering
distribution function which describes how much of the light coming
from direction {right arrow over (w)}' is reflected, refracted or
otherwise scattered in the direction {right arrow over (w)}, and is
generally the sum of a diffuse component, a glossy component and a
specular component, in Equation (1.22), the function G ({right
arrow over (x)}, {right arrow over (x)}') is a geometric term
G ( x .fwdarw. , x .fwdarw. ' ) = cos .theta.cos .theta. ' x
.fwdarw. - x .fwdarw. ' ( 1.23 ) ##EQU00027##
where .theta. and .theta.' are angles relative to the normals of
the respective surfaces at points {right arrow over (x)} and {right
arrow over (x)}' respectively. Further in Equation (1.22), V({right
arrow over (x)}, {right arrow over (x)}') is a visibility function
which equals the value one if the point {right arrow over (x)}' is
visible from the point {right arrow over (x)} and zero if the point
{right arrow over (x)}' is not visible from the point {right arrow
over (x)}.
[0100] The computer graphics system 10 makes use of global
illumination specifically in connection with determination of the
diffuse component
T.sub.f.sub.dT.sub.f.sub.dL
and the caustic component
T.sub.f.sub.dT.sub.f.sub.g.sub.+f.sub.gL
using a photon map technique. Generally, a photon map is
constructed by simulating the emission of photons by light
source(s) in the scene and tracing the path of each of the photons.
For each simulated photon that strikes a surface of an object in
the scene, information concerning the simulated photon is stored in
a data structure referred to as a photon map, including, for
example, the simulated photon's color, position, and direction
angle. Thereafter a Russian roulette operation is performed to
determine the photon's state, i.e., whether the simulated photon
will be deemed to have been absorbed or reflected by the surface,
if a simulated photon is deemed to have been reflected by the
surface, the simulated photon's direction is determined using, for
example, a bidirectional reflectance distribution function
("BRDF"). If the reflected simulated photon strikes another
surface, these operations will be repeated (see Grabenstein). The
data structure in which information for the various simulated
photons is stored may have any convenient form; typically
k-dimensional trees, for k an integer, are used. After the photon
map has been generated, it can be used in rendering the respective
components of the image.
[0101] In generating a photon map, the computer graphics system 10
simulates photon trajectories, thus avoiding the necessity of
discretizing the kernel of the underlying integral equation. The
interactions of the photons with the scene, as described above, are
stored and used for density estimation. The computer graphics
system 10 makes use of a scrambled low-discrepancy
strictly-deterministic sequence, such as a scrambled Halton
sequence, which has better discrepancy properties in higher
dimensions than does an unscrambled sequence. The scrambled
sequence also has the benefit, over a random sequence, that the
approximation error decreases more smoothly, which will allow for
use of an adaptive termination scheme during generation of the
estimate of the integral. In addition, since the scrambled sequence
is strictly deterministic, generation of estimates can readily be
parallelized by assigning certain segments of the low-discrepancy
sequence to ones of a plurality of processors, which can operate on
portions of the compulation independently and in parallel. Since
usually the space in which photons will be shot, by selecting
directions will be much larger than the area of the light sources
from which the photons were initially shot, it is advantageous to
make use of components of smaller discrepancy, for example, or
.PHI..sub.2 or .PHI..sub.3 (where, as above, .PHI..sub.b refers to
the radical inverse function for base b), for use in connection
with angles at which photons are shot, and components of higher
discrepancy, for example, scrambled .PHI..sub.5 or .PHI..sub.7, for
use in connection with sampling of the area of the respective light
source, which will facilitate filling the space more uniformly.
[0102] The computer graphics system if) estimates the radiance from
the photons in accordance with
L .fwdarw. r ( x , .omega. ) .apprxeq. 1 A i .di-elect cons. B k (
x ) f r ( .omega. i , x , .omega. ) .PHI. i ( 1.24 )
##EQU00028##
where, in Equation (1.24), .PHI..sub.i represents the energy of the
respective i-th photon, .omega..sub.i is the direction of incidence
of the i-th photon, B.sub.k(x) represents the set of the k nearest
photons around the point x, and A represents an area around point x
that includes the photons in the set B.sub.k(x). Since the energy
of a photon is a function of its wavelength, the .PHI..sub.i in
Equation (1.24) also represents the color of the respective i-th
photon. The computer graphics system 10 makes use of an unbiased
but consistent estimator for the area A for use in Equation (1.24),
which is determined as follows. Given a query ball, that is, a
sphere that is centered at point x and whose radius r (B.sub.k(x))
is the minimal radius necessary for the sphere to include the
entire set B.sub.k(x), a tangential disk D of radius r (B.sub.k(x))
centered on the point x is divided into M equal-sized subdomains
D.sub.i, that is
i = 0 M - 1 D i = D and D i D j .noteq. 0 for i .noteq. j , where D
i = D M = .pi. r 2 ( B k ( x ) ) M ( 1.25 ) ##EQU00029##
[0103] The set
P={D.sub.i|D.sub.i.andgate.{x.sub.i|.sub.D|i.di-elect
cons.B.sub.k(x)}.noteq.0} (1.26)
contains all the subdomains D.sub.i that contain a point
x.sub.i|.sub.D on the disk, which is the position of the i-th
photon projected onto the plane defined by the disk D along its
direction of incidence .omega..sub.i. Preferably, the number M of
subdomains will be on the order of {square root over (k)} and the
angular subdivision will be finer than the radial subdivision in
order to capture geometry borders. The actual area A is then
determined by
A = .pi. r 2 ( B k ( x ) ) P M ( 1.27 ) ##EQU00030##
[0104] Determining the actual coverage of the disk D by photons
significantly improves the radiance estimate in Equation (1.24) in
corners and on borders, where the area obtained by the standard
estimate .pi.r.sup.2(B.sub.k(x)) would be too small, which would be
the case at corners, or too large, which would be the case at
borders. In order to avoid blurring of sharp contours of caustics
and shadows, the computer graphics system 10 sets the radiance
estimate L to black if all domains D.sub.i that touch point x do
not contain any photons.
[0105] It will be appreciated that, in regions of high photon
density, the k nearest photons may lead to a radius r(B.sub.k(x))
that is nearly zero, which can cause an over-modulation of the
estimate. Over-modulation can be avoided by selecting a minimum
radius r.sub.min, which will be used if r (B.sub.k(x)) is less than
r.sub.min. In that, case, instead of Equation (1.24), the estimate
is generated in accordance with the following equation:
L _ r ( x , .omega. ) = N k i .di-elect cons. B k ( x ) .PHI. i f r
( .omega. i , x , .omega. r ) ( 1.28 ) ##EQU00031##
assuming each photon is started with 1/N of the total flux .PHI..
The estimator in Equation (1.28) provides an estimate for the mean
flux of the k photons if r(B.sub.k(x))<r.sub.min.
[0106] The global photon map is generally rather coarse and, as a
result, subpixel samples can result in identical photon map
queries. As a result, the direct visualization of the global photon
map is blurry and it is advantageous to perform a smoothing
operation in connection therewith hi performing such an operation,
the computer graphics system 10 performs a local pass integration
that removes artifacts of the direct visualization. Accordingly,
the computer graphics system 10 generates an approximation for the
diffuse illumination term T.sub.f.sub.gT.sub.f.sub.gL as
T f d T f d L .apprxeq. ( T f d L _ r ) ( x ) = .intg. S 2 ( x ) f
d ( x ) L _ r ( h ( x , .omega. .fwdarw. ) ) cos .theta. .omega.
.apprxeq. f d ( x ) M i = 0 M - 1 L _ r ( h ( x , .omega. ( arcsin
u i , 1 , 2 .pi. u i , 2 ) ) ) ( 1.29 ) ##EQU00032##
[0107] with the integral over the hemisphere S.sup.2(x) of incident
directions aligned by the surface normal in x being evaluated using
importance sampling. The computer graphics system 10 stratifies the
sample points on a two-dimensional grid by applying dependent
trajectory splitting with the Hammersley sequence and thereafter
applies irradiance interpolation. Instead of storing the incident
flux .PHI..sub.i of the respective photons, the computer graphics
system 10 stores their reflected diffuse power
f.sub.d(x.sub.i).PHI..sub.i with the respective photons in the
photon map, which allows for a more exact approximation than can be
obtained by only sampling the diffuse BRDF in the hit points of the
final gather rays. In addition, the BRDF evaluation is needed only
once per photon, saving the evaluations during the final gathering.
Instead of sampling the full grid, the computer graphics system 10
uses adaptive sampling, in which refinement is triggered by
contrast, distance traveled by the final gather rays in order to
more evenly sample the projected solid angle, aid the number of
photons that are incident form the portion of the projected
hemisphere. The computer graphics system 10 fills in positions in
the grid that are not sampled by interpolation. The resulting image
matrix of the projected hemisphere is median filtered in order to
remove weak singularities, after which the approximation is
generated. The computer graphics system 10 performs the same
operation in connection with, for example, hemispherical sky
illumination, spherical high dynamic-range environmental maps, or
any other environmental light source.
[0108] The computer graphics system 10 processes final gather rays
that strike objects that do not cause caustics, such as plane glass
windows, by recursive ray tracing. If the hit point of a final
gather ray is closer to its origin than a predetermined threshold,
the computer graphics system 10 also performs recursive ray
tracing. This reduces the likelihood that: blurry artifacts will
appear in corners, which might otherwise occur since for close hit
points the same photons would be collected, which can indirectly
expose the blurry structure of the global photon map.
[0109] Generally, photon maps have been taken as a snapshot at one
point in time, and thus were unsuitable in connection with
rendering of motion blur. Following the observation that averaging
the result of a plurality of photon maps is generally similar to
querying one photon map with the total number of photons from all
of the plurality of photon maps, the computer graphics system 10
generates N.sub.T photon maps, where N.sub.T is determined as
described above, at points in time
t b = t 0 + b + 1 2 N T T ( 1.30 ) ##EQU00033##
0.ltoreq.b<N.sub.T. As noted above, N.sub.T can equal "one," in
which case N photon maps are used, with N being chosen as described
above. In that case,
t.sub.i=t.sub.0+.PHI..sub.3(i)T (1.31)
and thus t.sub.i,j=t.sub.i,0, that is, t.sub.i, for N.sub.T=1. In
the general case (Equation (1.30)), during rendering, the computer
graphics system 10 uses the photon map with the smallest time
difference |t.sub.i,j-t.sub.b| in connection with rendering for the
time sample point t.sub.i,j.
[0110] The invention provides a number of advantages. In
particular, the invention provides a computer graphics system that
makes use of strictly deterministic distributed ray tracing based
on low-discrepancy sampling and dependent trajectory splitting in
connection with rendering of an image of a scene. Generally,
strictly deterministic distributed ray tracing based on
deterministic low-discrepancy sampling and dependent trajectory
splitting is simpler to implement than an implementation based on
random or pseudo-random numbers. Due to the properties of the
radical inverse function, stratification of sample points is
intrinsic and does not need to be considered independently of the
generation of the positions of the sample points. In addition,
since the methodology is strictly deterministic, it can be readily
parallelized by dividing the image into a plurality of tasks, which
cat be executed by a plurality of processors in parallel. There is
no need to take a step of ensuring that positions of sample points
are not correlated, which is generally necessary if a methodology
based on random or pseudo-random numbers is to be implemented for
processing in parallel.
[0111] Moreover, the methodology can be readily implemented in
hardware, such as a graphics accelerator, particularly if
Hammersley point sets are used, since all points with a fixed index
i yield a regular grid. A graphics accelerator can render a
plurality of partial images corresponding to these regular grids in
a number of parallel tasks, and interleave the partial images in an
accumulation buffer to provide the final image. Operating in this
manner provides very good load balancing among the parallel tasks,
since all of the tasks render almost the same image.
[0112] In addition, the methodology can readily be extended to
facilitate rendering of animations. Generally, an animation
consists of a series of frames, each frame comprising an image. In
order to decorrelate the various frames, instead of initializing
the field of integers used as identifiers for ray instances for
each frame by i, i+i.sub.f can be used, where i.sub.f is a frame
number. This operates as an offsetting of i by i.sub.f, which is
simply a generalization of the Hammersley points. A user can select
to initialize the field of integers for each frame by i, in which
case the frames will not be correlated. In that case, undersampling
artifacts caused by smooth motion will remain local and are only
smoothly varying. Alternatively, the user can select to initialize
the field of integers for each frame by i+i.sub.f, in which case
the artifacts will not remain local, and will instead appear as
noise or film grain flicker in the final animation. The latter is
sometimes a desired feature of the resulting animation, whether for
artistic reasons or to match actual, film grain. Another variation
is to add i.sub.f directly to k and clip the result by 2.sup.n
(reference Code Segment 1, below). In that case, the pixel sampling
pattern will change from frame to frame and the frame number
i.sub.f will need to be known in the post-production process in
order to reconstruct the pixel sampling pattern for compositing
purposes.
[0113] Generally, a computer graphics system that makes use of
deterministic low-discrepancy sampling in determination of sample
points will perform better than a computer graphics system that
makes use of random or pseudo-random sampling, but the performance
may degrade to that of a system that makes use of random or
pseudo-random sampling in higher dimensions. By providing that the
computer graphics system performs dependent splitting by
replication, the superior convergence of low-dimensional
low-discrepancy sampling can be exploited with the effect that the
overall integrand becomes smoother, resulting in better convergence
than with stratified random or pseudo-random sampling. Since the
computer graphics system also makes use of dependent trajectory
sampling by means of infinite low discrepancy sequences, consistent
adaptive sampling of, for example, light sources, can also be
performed.
[0114] In addition, it will be appreciated that, although the
computer graphics system has been described as making use of sample
points generated using generalized scrambled and/or unscrambled
Hammersley and Halton sequences, it will be appreciated that
generally any (t, m, s)-net or (t, s)-sequence can be used.
[0115] At a more general level, the invention provides an improved
quasi-Monte Carlo methodology for evaluating an integral of a
function f on the s-dimensional unit cube [0, 1).sup.s. In contrast
with this methodology, which will be referred to as trajectory
splitting by dependent splitting, in prior methodologies, fee
sample points in the integration domain for which the sample values
at which sample values for the function were generated were
determined by providing the same number of coordinate samples along
each dimension. However, for some dimensions of an integrand, it is
often the case that the function f will exhibit a higher variance
than for other dimensions. The invention exploits this by making
use of trajectory splitting by dependent samples in critical
regions.
[0116] The partial integral
g ( x ) = .intg. I s 2 f ( x , y ) y .apprxeq. 1 N 2 j = 0 N 2 - 1
f ( x , y j ) ( 1.32 ) ##EQU00034##
(x and y comprising disjoint sets of the s-dimensions, and
x.orgate.y comprising the set of all of the dimensions), where
N.sub.2 identifies the number of samples selected for the set y of
dimensions, can be defined over the portion of the integration
domain that is defined by unit cube (0, 1].sup.s.sup.2, which, in
turn, corresponds to the portion of the integration domain that is
associated with set s.sub.2 dimensions. Evaluating g(x) using
Equation (1.32) will affect a smoothing of the function f in the
s.sub.2 dimensions that are associated with set y.
[0117] The result generated by applying Equation (1.32) can then be
used to evaluate the full integral
.intg. I s 1 .intg. I s 2 f ( x , y ) y x = .intg. I s 1 g ( x ) x
.apprxeq. 1 N 1 i = 0 N 1 - 1 1 N 2 j = 0 N 2 - 1 f ( x i , y j ) (
1.33 ) ##EQU00035##
where N.sub.1 identifies the number of samples selected for the set
x of dimensions, that is, over the remaining dimensions of the
integration domain that are associated with the dimensions that are
associated with the set x. If the dimension splitting x,y is
selected such that the function f exhibits relatively high variance
over the set y of the integration domain, and relatively low
variance over the set x, it will not be necessary to generate
sample values for the function N.sub.1-times-N.sub.2 times. In that
case, it will suffice to generate sample only values N.sub.2 times
over the integration domain. If the correlation coefficient of
f(.xi., .eta.) and f(.xi., .eta.'), which indicates the degree of
correlation between values of function evaluated, for the former,
at (x.sub.i, y.sub.i)=(.xi., .eta.), and, for the later, at b
(x.sub.i, y.sub.i)=(.xi., .eta.'), is relatively high, the time
complexity required to evaluate the function
f.sub.[0,1)g(x.sub.0, . . . , x.sub.s-1)
will be decreased.
[0118] The smoothness of an integrand can be exploited using a
methodology that will be referred to as correlated sampling.
Generally, that is, if correlated sampling is not used, in
evaluating an integral each dimension will be associated with its
respective sequence. However, in correlated sampling, the same
sequence can be used for all of the dimensions over the integration
domain, that is
1 M j - 1 N 1 N j i = 0 N j - 1 f j ( x i , y j ) .apprxeq. 1 M j =
1 M .intg. I s f j ( x ) x = .intg. I s 1 M j = 1 M f j ( x ) x
.apprxeq. 1 N i = 0 N - 1 1 M j = 1 M f j ( x i ) ( 1.34 )
##EQU00036##
[0119] The methodology of trajectory splitting by depending
sampling makes use of a combination of the trajectory splitting
technique described above in connection with Equations (1.32) and
(1.33) with the correlated sampling methodology described in
connection with Equation (1.34).
[0120] Since integrals are invariant under toroidal shifting for
z.sub.j.di-elect cons.I.sup.s.sup.2, that is,
S j : I s 2 .fwdarw. I s 2 y ( y + z j ) mod 1 .intg. I s 2 g ( y )
y = .intg. I s 2 g ( S j ( y ) ) y ( 1.35 ) ##EQU00037##
the values of the integrals also do not change. Thus, if, in
Equation (1.33), the inner integral is replicated M times,
.intg. I s 1 .intg. I s 2 f ( x , y ) y x = .intg. I s 1 .intg. I s
2 1 M j = 0 M - 1 f ( x , S j ( y ) ) y x .apprxeq. 1 N i = 0 N - 1
1 M j = 0 M - 1 f ( x i , S j ( y i ) ) = 1 N i = 0 N - 1 1 M j = 0
M - 1 f ( x i , ( y i + z j ) mod 1 ) ( 1.36 ) ##EQU00038##
[0121] For index j, the functions f(x.sub.i, S.sub.j(y.sub.i)) are
correlated, enabling the smoothness of the integrand in those
dimensions that are represented by y to be exploited, as
illustrated above in connection with Equation (1.19) (lens
sampling). Equations (1.20) and (1.2.1) (area light sources) and
Equation (1.29) (approximation for fee diffuse illumination term).
It will be appreciated that the evaluation using the replication is
the repeated application of fee local quadrature rule
U.sub.M,s.sub.g:=(z.sub.j).sub.j=0.sup.M
shifted by random offset values y.sub.i. The use of dependent
variables in this manner pays off particularly if there is some
smoothness in the integrand along one or more dimensions. Splitting
can be applied recursively, which yields a history tree, in which
each path through the respective history tree represents a
trajectory of a particle such as a photon.
[0122] The quasi-Monte Carlo methodology of trajectory splitting by
dependent sampling makes use of sets of deterministic,
low-discrepancy sample points both for the global quadrature
rule
U.sub.N,s.sub.1.sub.+s.sub.2=(x.sub.i,y.sub.i).sub.i=0.sup.N
that is, integration over all of the dimensions s.sub.1+s.sub.2
comprising the entire integration domain, as well as for the local
quadrature rule
U.sub.M,s.sub.g:=(z.sub.j).sub.j=0.sup.M
that is, integration over the dimensions s.sub.2 of the integration
domain. The methodology unites splitting and dependent sampling,
exploiting the stratification properties of low-discrepancy
sampling. Accordingly, it will be possible to concentrate more
samples along those dimensions in which the integrand exhibits high
levels of variation, and fewer samples along those dimensions in
which the integrand exhibits low levels of variation, which reduces
the number of sample points at which the function will need to be
evaluated. If the methodology is to be applied recursively a
plurality of times, it will generally be worthwhile to calculate a
series of values z.sub.j that are to comprise the set
U.sub.M,s.sub.2. In addition, the methodology may be used along
with importance sampling and, if U is an infinite sequence,
adaptive sampling. In connection with adaptive sampling, the
adaptations will be applied in the replication independently of the
sampling rate, so that the algorithm will remain consistent. The
low-discrepancy sample points sets U.sub.N,s.sub.1.sub.+s.sub.2 and
U.sub.M,s.sub.2 can be chosen arbitrarily; for example, the sample
point set U.sub.M,s.sub.2 can be a projection of sample point set
U.sub.N,s.sub.1.sub.+s.sub.2. When trajectory splitting is
recursively applied to build trajectory trees, generalizing the
point set U.sub.N,s.sub.1.sub.+s.sub.2 for the subsequent branches
can be used to decorrelate the separate parts of the respective
tree.
[0123] FIG. 5 shows a code fragment 140, referred to herein as
"Code Segment 1," in the C++ programming language for generating
the positions of the jittered subpixel sample points x.sub.i. FIG.
6 shows a code fragment 142, referred to herein as "Code Segment
2," in the C++ programming language for generating a ray tree class
Ray.
[0124] It will be appreciated that a system in accordance with the
invention can be constructed in whole or in part from special
purpose hardware or a general purpose computer system, or any
combination thereof, any portion of which may be controlled by a
suitable program. Any program may in whole or in part comprise part
of or be stored on the system in a conventional manner, or it may
in whole or in part be provided in to the system over a network or
other mechanism for transferring information in a conventional
manner. In addition, it will be appreciated that the system may be
operated and/or otherwise controlled by means of information
provided by an operator using operator input elements (not shown)
which may be connected directly to the system or which may transfer
the information to the system over a network or other mechanism for
transferring information in a conventional manner.
[0125] With these points in mind, we next turn to image synthesis
by adaptive quasi-Monte Carlo integration.
II. Image Synthesis by Adaptive Quasi-Monte Carlo Integration
[0126] Analyzing the implicit stratification properties of the
deterministically scrambled Halton sequence leads to an adaptive
interleaved sampling scheme that improves many rendering
algorithms. Compared to uncorrelated adaptive random sampling
schemes, the correlated and highly uniform sample points from, the
incremental Halton sequence result in a faster convergence and much
more robust adaptation. Since fee scheme is deterministic,
parallelization and reproducibility become trivial, while
interleaving maximally avoids aliasing. The sampling scheme
described herein are useful in a number of applications, including,
for example, industrial path tracing, distribution ray tracing, and
high resolution compositing.
[0127] As discussed above, the process of image synthesis includes
computing the color of each pixel in the image. The pixel color
itself is determined by an integral. Due to high dimensionality and
unknown discontinuities of the integrand, this integral typically
must be approximated using a numerical technique, such as the Monte
Carlo method. The efficiency of the image synthesis process can be
significantly improved by using adaptive schemes that take into
account variations in the complexity of the integrands.
[0128] Analytical integration methods developed for computer
graphics perform very-well for small problems, i.e., low integrand
dimension or untextured scenes. Approaches like discontinuity
meshing or approximate analytic integration, however, utterly fail
for, e.g., higher order shadow effects. Consequently, high-end
rendering algorithms rely on sampling.
[0129] Starting from early computer graphics, many adaptive
sampling schemes have been developed in order to control rendering
effort. A number of these schemes rely on partitioning the
integration domain along each axis. As long as only low-dimensional
integration (e.g., pixel anti-aliasing) was of interest, the
inherent curse of dimension of axis-aligned recursive refinement
was not perceptible, (The term "curse of dimension" refers to the
known issue that, computational cost, typically increases
exponentially with the dimension of a problem.) These schemes are
still applied today. However, due to the curse of dimension, for
example, in distribution ray tracing and global illumination
simulation, these schemes are applied only to the lowest dimensions
of the integrands, e.g., to the image plane.
[0130] Many adaptive schemes rely on comparing single samples in
order to control refinement. For example edges are detected by
comparing contrast against a threshold. Such schemes fail in two
aspects: If the contrast does not indicate refinement, nevertheless
important contributions to the image can be missed. On the other
hand refinement can be taken too far. This happens, for example,
when sampling an infinite black-and-white checkerboard in
perspective. At the horizon, refinement is driven to full depth,
although the correct gray pixel color may already be obtained by
one black and white sample.
[0131] In fact, the paradigm of sample-based refinement performs
function approximation of the integrand itself although only
averages of the integrand are required. This is considered by
pixel-selective Monte Carlo schemes that consider the estimated
variance of the estimate of the functional to be computed. However,
Monte Carlo error estimation requires independent random samples,
which limits the amount of uniformity of the samples and thus
convergence speed.
[0132] Thinking of adaptive sampling as image processing it is easy
to identify noise or edges in an image by computing derivatives
between pixels in order to trigger refinement.
[0133] By considering image synthesis as computing families of
functionals instead of independent pixel values only, a powerful
adaptive sampling scheme has been developed. Therefore stratified
sequences of sample points are extracted from the scrambled Halton
sequence. Although these points are deterministic, aliasing is
avoided maximally. Incorporating tone mapping directly into the
pixel integral, in addition to high uniformity of the subsequences,
yields a superior and smooth convergence. Consequently adaptation
can be controlled robustly by applying simple image processing
operators to the final pixel values rather than to single samples.
Since everything is deterministic exact reproducibility as required
in production is trivial. The superior performance of the new
technique is described below with respect to various
applications.
[0134] The scrambled Halton sequence is now described. For the
purposes of the present discussion, filtering, tone mapping, and
actual radiance computations are hidden in the integrand f defined
on the s-dimensional unit cube. The color of a pixel is then
determined by an integral
.intg. [ 0 , 1 ) s f ( x ) x .apprxeq. 1 N j = 0 N - 1 f ( x j )
##EQU00039##
which is numerically is approximated by averaging N function
samples at the positions x.sub.j.di-elect cons.[0,1).sup.s.
[0135] It has been demonstrated non-adaptive quasi-Monte Carlo
integration is highly efficient in computer graphics. The used
so-called quasi-Monte Carlo points are of low discrepancy, meaning
that due to high correlation they are much more uniformly
distributed than random samples can be. Due to their deterministic
nature, however, unbiased error estimation from the samples
themselves is not possible opposite to the Monte Carlo method.
[0136] In order to take advantage of the much faster and smoother
convergence of quasi-Monte Carlo integration and to obtain a
reliable adaptation control, important properties of the scrambled
Halton sequence are described in the following discussion. This
deterministic low discrepancy point is easily constructed as can be
seen in the sample code, discussed below.
[0137] "Stratification by radical inversion" is now described. The
radical inverse
.PHI. b : 0 -> [ 0 , 1 ) i = l = 0 .infin. .alpha. l ( i ) b l l
= 0 .infin. .alpha. l ( i ) b - l - 1 ( 2.1 ) ##EQU00040##
mirrors the representation of the index i by the digits
.alpha..sub.g(i).di-elect cons.{0, . . . , b-1}
in the integer base b.di-elect cons.N at the decimal point. In base
b=2, this means that
.PHI. 2 ( i ) .di-elect cons. { [ 0 , 1 / 2 ) for i even , [ 1 / 2
, 1 ) else . ( 2.2 ) ##EQU00041##
This observation has been generalized, and given the name
periodicity. In fact, however, this property relates much more to
stratification as formalized by the definition of (0, 1)-sequences.
Therefore, a different derivation is presented herein that stresses
the actual stratification property. The index i is chosen as
i=jb.sup.n+k for k.di-elect cons.{0, . . . , b.sup.n-1}
and inserted into Equation (2.1) yielding
.PHI. b ( i ) = .PHI. b ( j b n + k ) = l = 0 .infin. a l ( j b n +
k ) b - l - 1 = l = 0 .infin. a l ( j b n ) b - l - 1 + l = 0
.infin. a l ( k ) b - l - 1 = l = 0 .infin. a l ( j ) b - n - l - 1
+ .PHI. b ( k ) = b - n .PHI. b ( j ) .di-elect cons. [ 0 , b - n )
+ .PHI. b ( k ) ( 2.3 ) ##EQU00042##
by exploiting the additivity of the radical inverses that belong to
the class of (0, 1)-sequences. The first term of the result depends
on/and obviously is bounded by b.sup.-n, while the second term is a
constant offset by the radical inverse of k. Since
k.di-elect cons.{0, . . . , b.sup.n-1}
it follows that
.PHI..sub.k.di-elect cons.{0,b.sup.-n,2b.sup.-n, . . . ,
(b.sup.n-1)b.sup.-n}
Fixing the n first digits by k then for j.di-elect cons.N.sub.0 2
yields radical inverses only in the interval
[.PHI..sub.b(k),.PHI..sub.b(k)+b.sup.-n)
[0138] There are now described stratified samples from the Halton
sequence. For quasi-Monte Carlo integration, multidimensional
uniform samples are required. However, from Equation (2.2) it may
be seen that the radical inverse is not completely uniformly
distributed, i.e., cannot just replace the random number generator.
Therefore, multidimensional uniform deterministic samples may be
constructed, for example, by the Halton sequence
x.sub.i=(.PHI..sub.b.sub.g(i), . . . ,
.PHI..sub.b.sub.g(i)).di-elect cons.[0,1).sup.g
where for the c-th component b.sub.c is the c-th prime number.
[0139] The one-dimensional observations above generalize to higher
dimensions. Stratified sampling is implicitly embodied. This is
seen by choosing the index
i = j d = 1 s b d n d + k with 0 .ltoreq. k < d = 1 s b d n d (
2.4 ) ##EQU00043##
yielding
.PHI. b c ( i ) = .PHI. b c ( j d = 1 s b d n d + k ) = b c - n c
.PHI. b c ( j d = 1 d .noteq. c s b d n d ) + .PHI. b c ( k )
##EQU00044##
analogous to Equation (2.3), and consequently
x i .di-elect cons. c = 1 s [ .PHI. b c ( k ) , .PHI. b c ( k ) + b
c - n c ) ##EQU00045##
for fixed k and j.di-elect cons..sub.0 with the choice of the index
i according to Equation (2.4). It may be seen that the
.PI..sub.d=1.sup.sb.sub.d.sup.n.sup.d
disjoint strata selected by k are disjoint and form a partition of
the unit cube [0, 1).sup.s. However, the scheme suffers the curse
of dimension, since stratifying all s dimensions results in an
exponential growth of the number of strata. This may be the reason,
why only the first four dimensions have been stratified.
[0140] The stratification property is illustrated in FIGS. 7A and
7B, which show plots 200 and 210, respectively, of the first two
components x.sub.i=(.PHI..sub.2(i), .PHI..sub.3(i)) of the Halton
sequence for 0.ltoreq.i<j6+k<2.sup.43.sup.3=216. The stratum
with the emphasized points contains all indices i with k=1. Scaling
the strata to be square, i.e.,
x.sub.i.fwdarw.x.sub.i.sup.g=(2.sup.1.PHI..sub.2(i),3.sup.1.PHI..sub.3(i-
))
does not affect discrepancy, since it is a separable mapping.
[0141] Deterministic scrambling is now described. The Halton
sequence exposes superior uniformity properties, i.e., low
discrepancy and minimum distance property. However, low-dimensional
projections exhibit correlation patterns. FIGS. 8A and 8B show a
pair of low-dimensional projections 220 and 230. FIG. 8A shows the
Halton sequence for the points
(.PHI..sub.17(i),.PHI..sub.19(i)).sub.i=0.sup.255
FIG. 8B shows the scrambled Halton sequence for the points
(.PHI.'.sub.17(i),.PHI.'.sub.19(i)).sub.i=0.sup.255
As illustrated by FIGS. 8A and 8B, scrambling can significantly
improve uniformity.
[0142] While usually not perceptible, this low-dimensional
correlation often interferes with the integrands in computer
graphics that have low-dimensional structure. For example,
correlation can slow down convergence in a sequence of
two-dimensional scattering events as used in path tracing.
[0143] One remedy is to scramble the Halton sequence. The radical
inverse is replaced by the scrambled radical inverse
.PHI. b ' : 0 -> [ 0 , 1 ) i = l = 0 .infin. .alpha. l ( i ) b l
l = 0 .infin. .pi. b ( a l ( i ) ) b - l - 1 ( 2.5 )
##EQU00046##
yielding a scrambled Halton sequence
x'.sub.i:=(.PHI.'.sub.b.sub.1(i), . . . ,
.PHI.'.sub.b.sub.s(i))
The scrambling permutations .pi..sub..delta. applied to the digits
a.sub.g(i) are determined by a recursion starting with the identity
.pi..sub.2=(0, 1) If b is odd, .pi..sub.b is constructed from
.pi..sub.b-1 by incrementing each value
.gtoreq. b - 1 2 ##EQU00047##
and inserting
b - 1 2 ##EQU00048##
in the middle. Otherwise, .pi..sub.b is constructed from
.pi..sub.b-1 by concatenating
2 .pi. b 2 ##EQU00049## and ##EQU00049.2## 2 .pi. b 2 + 1
##EQU00049.3##
This algorithm yields
.pi..sub.2=(0,1)
.pi..sub.3=(0,1,2)
.pi..sub.4=(0,2,1,3)
.pi..sub.5=(0,3,2,1,4)
.pi..sub.6=(0,2,4,1,3,5)
.pi..sub.7=(0,2,5,3,1,4,6)
.pi..sub.8=(0,4,2,6,1,5,3,7)
The scrambling improves the uniformity. This is especially visible
for low-dimensional projections as illustrated in FIGS. 8A and 8B.
In addition, the minimum distance of samples is increased, which
indicates an increased uniformity. An implementation is described
below.
[0144] The observations from Equation (2.3) and from the above
discussion transfer to the scrambled Halton sequence in a
straightforward way. This can be seen for two-dimensional
stratification, since .pi..sub.2 and .pi..sub.3 are identities and
consequently
.PHI.'.sub.2.ident..PHI..sub.2
and
.PHI.'.sub.3.ident..PHI..sub.3
[0145] There is now described a technique for bias elimination by
randomization. By construction, radical inversion generates only
rational numbers .andgate.[0, 1) as set forth in Equations (2.1)
and (2.5). Nevertheless, it can be shown that quasi-Monte Carlo
integration is biased but consistent for Riemann-integrable
functions.
[0146] If required, the bias can be removed by randomly shifting
the deterministic points
x'.sub.i.fwdarw.x'.sub.i+.xi. mod
1=(.PHI..sub.b.sub.1(i)+.xi..sup.(g)mod 1, . . . ,
.PHI.'.sub.b.sub.2(i)+.di-elect cons..sup.(s)mod 1)
of the scrambled Halton sequence modulo one, where
.xi.=(.xi..sup.(g), . . . , .xi..sup.(s)).di-elect
cons.[0,1).sup.s
is a vector of s independent realizations of uniform random numbers
on the unit interval. The resulting minimally randomized estimator
has been analyzed, and a variance reduction of
.sigma. 2 ( 1 N i = 0 N - 1 f ( x i ' + .xi. mod 1 ) ) = ( ln 2 s N
N 2 ) ##EQU00050##
has been proofed for square integrable functions. Multiple
realizations of this and other randomization techniques allow the
estimation of the variance in order to control the integration
error. Since the error is controlled in a different way, one
instance of such a randomization is sufficient to cancel the bias.
However, in fact, the bias of the scrambled Halton sequence
compared to the randomly shifted points, and consequently the
randomization becomes negligible.
[0147] Considering random scrambling leads to a second important
observation. Often the uniformity of the samples is improved;
however, some realizations also can decrease the uniformity of a
point set by, e.g., lowering the mutual minimum distance. Our
experiments show that random scrambling only marginally changes the
uniformity of the scrambled Halton sequence, indicating that the
deterministic permutations .pi..sub.b, which are themselves a
subset of the permutations available for random scrambling, already
are a very good choice. In addition implementing the scrambled
Halton sequence, as described below, is simpler than random
scrambling.
[0148] From the above observations, it can be concluded that
randomization in not necessary for the presently described
applications. The structure of the deterministic scrambled Halton
sequence is described below with respect to specific
implementations.
[0149] A new technique for image synthesis is now described. Image
synthesis includes computing a matrix of pixel colors
I m , n := .intg. [ 0 , 1 ) s f m , n ( x ) x .apprxeq. j = 0 N m ,
n - 1 w j , m , n R .alpha. ( L ( x j , m , n ) ) ( 2.6 )
##EQU00051##
The function f.sub.m, n to be integrated for the pixel at position
(m, n) usually contains discontinuities and can be of high
dimension. Efficient general analytical solutions are not available
and consequently the integrals have to be numerically approximated.
In addition the complexity of the integrands varies so that
adaptive integration pays off.
[0150] Deterministic anti-aliasing is now described. The first
component x.sub.i.sup.(1) is scaled by b.sub.1.sup.n.sup.1 and the
second x.sub.i.sup.(2) by b.sub.2.sup.n.sup.2 as illustrated in
FIGS. 7A and 78. Thus, a
b.sub.1.sup.n.sup.1.times.b.sub.2.sup.n.sup.2 stratified sample
pattern is obtained that is periodically tiled over the image
plane. FIG. 9 shows a plot 240 of the tiled sample pattern.
[0151] Identifying each stratum with a pixel, the identification k
is determined, for example by a table lookup, from the pixel
coordinates, and a Halton sequence restricted to that pixel is
obtained from
i=jb.sub.1.sup.n.sup.1b.sub.2.sup.n.sup.2+k for j.di-elect
cons..sub.0
[0152] This restriction means fixing the first n.sub.1 and n.sub.2
decimal digits of the first and second component, respectively, and
does not change the superior uniformity properties of the Halton
points. Consequently, the improved and smooth convergence of
deterministic low discrepancy sampling is preserved. Convergence is
improved substantially further by applying tone-mapping techniques
that in fact bound the integrands. Then, adaptation triggered by
image processing operators becomes very reliable. The number of
strata b.sub.1.sup.n.sup.1.times.b.sub.2.sup.n.sup.2 determined by
n.sub.1 and n.sub.2, which are chosen large enough, so that the
strata covered by adjacent pixel reconstruction filters do not
contain repeated patterns.
[0153] While for Monte Carlo integration it is easy to control
adaptation by estimating the integration error from the samples,
this is not possible for the correlated samples of quasi-Monte
Carlo integration.
[0154] A pixel is refined, whenever a refinement criterion is met.
As an example, a simple criterion indicates refinement by checking
the image gradient
.parallel..gradient.I.sub.m,n.parallel.<T.sub.m,n.sup.-.alpha.
against a threshold T. The exponent .alpha..di-elect cons.[0.5,
1.5] can be used to adapt to the speed of convergence. N.sub.m,n is
the number of samples used for pixel (m, n).
[0155] As known from music equipment, clipping signals causes
distortion. Therefore, instead of clipping, compression is used,
meaning that an upper bound on the signal is achieved by using a
continuously differentiable function. In addition, single signals
are compressed before being mixed.
[0156] According to an aspect of the invention, the same is done
for image synthesis. In Equation (2.6), the luminance L is
compressed by
R .alpha. : 0 + [ 0 , 1 ) ##EQU00052## L { L L < .alpha. .alpha.
+ ( 1 - .alpha. ) L - .alpha. 1 + L - .alpha. else
##EQU00052.2##
before averaging. By .alpha..di-elect cons.[0, 1], it is possible
to blend between linear transfer that is clipped at .alpha.=1 and
compression for .alpha.=0. The derivative of
x 1 + x ##EQU00053##
in x=0 is 1 and such the mapping R.sub..alpha. is continuously
differentiable. Depending on the output, media, many other response
curve mappings R.sub..alpha. are possible such as, for example, the
sRGB compression for video displays or film material
characteristics.
[0157] Performing the tone-mapping, i.e., compression, directly in
the quadrature is a remedy to the overmodulation problems from,
e.g., reflections or light source samples, and consequently
convergence is increased because the integrands are bounded now.
Thus, the noise level is reduced, contrary to the gradient domain
method and advanced filter kernel construction becomes
unnecessary.
[0158] In applying the above-described techniques to synthesize an
image, more samples are used to resolve certain details in the
final image. Adaptive samples are generated by refining the Halton
sequence in the first two dimensions by some contrast criterion in
screen space.
[0159] The above-described technique has a number of different
applications, including high-resolution compositing, frameless
rendering, and parallelization.
[0160] Based on the stratification properties of radical inversion
a new adaptive integration algorithm for image synthesis has been
presented. Using the stratification for interleaving, aliases are
efficiently suppressed, while the smooth convergence of
low-discrepancy sequences allows one to use a very efficient and
robust termination criterion for adaptation. Since all sample
positions are deterministic, storing the sampled function values
allows high-resolution compositing. In a similar way, the
convergence can be improved by using the microstructure of a (t,
s)-sequence, such as, for example, the Sobol sequence. While the
described techniques benefit from stratification in only two
dimensions, it is not suited for general high-dimensional
stratification due to the curse of dimension.
[0161] FIG. 10 shows a plot 250 illustrating an interleaved
adaptive supersampling technique, according to a further aspect of
the invention, which is tiled over the whole screen. The plotted
points are represented by:
(.PHI..sub.2(i),.PHI..sub.3(i))=P(i)
[0162] The image is scaled by (1, 1.5), and a 2.times.3
stratification is used. As shown in FIG. 10, the strata are
addressed by an offset, generating a point sequence in a subpixel.
The points are enumerated by the offset O of the desired stratum
plus the number of strata multiplied by the point number. In the
illustrated section of the plot, the point numbers are:
i=6j+3
i.di-elect cons.{0,1295}
[0163] In the equation i=6j+3, j is multiplied by 6 because of the
2.times.3 stratification. Also, i.di-elect cons.{0,
1295}=2.sup.43.sup.4.
[0164] Using the illustrated interleaving technique, adjacent
pixels are sampled differently, but strictly deterministically.
III. Additional Examples and Points Regarding Quasi-Monte Carlo
Integration
[0165] Computer graphics textbooks teach that sampling images using
deterministic patterns or lattices can result in aliasing. Aliasing
can only be avoided by random, i.e., independent sampling of
images. Thus, textbooks typically recommend random samples with
blue noise characteristic. However, these types of samples are
highly correlated due to their maximized minimum mutual distance.
Contrary to the textbook approach, the systems and techniques
described herein are based on parametric integration by quasi-Monte
Carlo methods, and are strictly deterministic.
[0166] Image synthesis is the most visible part of computer
graphics. One aspect, of image synthesis is concerned with the
synthesis of physically correct images. Thus, one image synthesis
technique includes identifying Sight paths that connect light
sources and cameras and summing up their respective contributions.
Another aspect of image synthesis is concerned with
non-photorealistic rendering including, for example, the simulation
of pen strokes or watercolor.
[0167] Image synthesis poses an integro-approximation problem for
which analytical solutions are available in exceptional cases only.
Therefore numerical techniques have to be applied. Prior art
approaches typically use elements of classical Monte Carlo
integration, in which random points are used to numerically
approximate the solution to an image integral. However, as
described herein, it is significantly more efficient to use
quasi-Monte Carlo integration, in which sequences of quasirandom
numbers are used to compute the solution to an image integral. The
presently described systems and techniques are useful, for example,
in a motion picture, which typically requires an extremely large
number of high-quality images.
[0168] The underlying mathematical task is to determine the
intensity I (k, l, t, .lamda.), where (k, l) is the location of a
pixel on the display medium. For the sake of clarity, there is
omitted the dependency on the time t and the wavelength .lamda. of
a color component of a pixel in the sequel.
[0169] Determining the intensity of a single pixel I (k, l), i.e.,
measuring the light flux through a pixel, requires to compute a
functional of the solution of the radiance transport integral
equation
L ( x , w ) = L e ( x , w ) + .intg. S 2 L ( h ( x , w i ) , - w i
) f ( w i , x , w ) cos .theta. i w i = : ( T f L ) ( x , w ) .
##EQU00054##
[0170] As a Fredholm integral equation of the second kind, the
radiance L in the point x into the direction .omega. is the sum of
the source radiance L.sub.e and the reflected and transmitted
radiance T.sub.fL, which is an integral over the unit sphere
S.sup.2. The cosine of the angle .theta..sub.i between the surface
normal in x and the direction of incidence .omega..sub.i accounts
for the perpendicular incident radiance only, which is colored by
the surface interface properties given by f. Finally h determines
the closest point of intersection of a ray from x into the
direction .omega..sub.i. The extension to participating media,
which we omit here for lack of space, exposes the same
structure.
[0171] Simultaneously computing all pixels
I(k,l):=.intg..sub..differential.V.intg..sub.S.sub.2R.sub..alpha.(L(x,.o-
mega.),k,l,x,.omega.)d.omega.dx
of an image is an integro-approximation problem. The mapping
R.sub..alpha. represents the mathematical description of the camera
and its response to the radiance L. R.sub..alpha. often is
non-linear in order to be able to compensate for the limited
dynamic range of most display media.
[0172] In a physically correct setting the norm
.parallel.T.sub.f.parallel. must be bounded by 1 in order to
guarantee energy conservation. Then the Neumann-series converges
and the computation of the radiance
L - SL c := i = 0 .infin. T f i L e ##EQU00055##
can be reduced to an infinite sum of integrals with increasing
dimension. The single integrals T.sub.f.sup.sL.sub.e have a
repetitive low dimensional structure inherited from stacking
transport operators. Obviously lower powers of the transport
operator are likely to be more important. Real world light sources
are bounded and consequently the radiance L uniformly can be
bounded by some b>0. In addition real world radiance
L ( y , w , t , .lamda. ) .di-elect cons. L b 2 ##EQU00056##
is a signal of finite energy and thus must be square
integrable.
[0173] However, often singular surface properties, as for example
specular reflection, are modeled by
(T.sub..delta..sub..omega.,L)(x,.omega.):=L(h(x,.omega.'),-.omega.')
using Dirac's .delta. distribution, where
.omega.'.ident..omega.'(.omega.) is the direction of specular
reflection. Then the operator norm of the solution operator can
even reach 1 and the Neumann series can diverge. The additional
problem of insufficient techniques is caused by
.delta.L.sub.b.sup.2
because some transport paths cannot be efficiently sampled and
force fee need of biased approximations like, e.g., the photon
mapping algorithm for rendering caustics.
[0174] Both the radiance L and the intensity I are non-negative and
piece wise continuous, where the discontinuities cannot be
efficiently predicted. The actual basis of the function class to
represent and approximate the intensity I (k, l, t, .lamda.) in
fact is determined by the display medium or image storage format,
e.g., an interleaved box basis for the color components of TFT
displays, cosines for JPEG compressed images, etc.
[0175] Due to the lack of efficient analytical solutions, rendering
algorithms reduce image synthesis to numerical
integro-approximation. Simulating a camera with anti-aliasing,
motion blur, and depth of field already contributes 5 dimensions to
fee integration domain of the intensity I. Area light sources and
each level, of reflection contribute another 2 dimensions.
Consequently the mathematical problem is high-dimensional,
discontinuous, and in L.sub.b.sup.2. Since tensor product
techniques will fail due to dimensionality and a Jack of
continuity, Monte Carlo and quasi-Monte Carlo methods are the
obvious choice.
[0176] Monte Carlo methods use random sampling for estimating
integrals by means. Quasi-Monte Carlo methods look like Monte Carlo
methods, however, they use deterministic points for sampling an
integrand. In contrast to random samples, the specifically designed
deterministic point sets are highly correlated, which allows for a
much higher uniformity and results in a faster convergence.
[0177] Real random numbers on the unit interval are characterized
by independence, unpredictability, and uniformity. For Monte Carlo
integration the independence is required to prove error bounds and
the uniformity is required to prove the order of convergence. Since
real random numbers are expensive to generate, usually efficient
deterministic algorithms are used to simulate pseudorandom numbers,
which then of course are perfectly predictable but seemingly
independent. However, the independence cannot be observed any
longer after averaging the samples.
[0178] Quasi-Monte Carlo integration is based on these
observations. By neglecting independence and unpredictability it is
possible to construct deterministic points, which are much more
uniform than random number samples can be. There exist a lot of
constructions for such deterministic point sets P.sub.n={x.sub.0, .
. . , x.sub.n-1}.OR right.[0, 1).sup.s, which are based on (1)
radical inversion based point sets; and (2) rank-1 lattice
points:
[0179] (1) Radical Inversion Based Point Sets Determine Samples
by
x i = ( i n , .PHI. b 1 ( i ) , , .PHI. b s - 1 ( i ) )
##EQU00057##
where
.PHI. b : 0 [ 0 , 1 ) i = l = 0 .infin. a l ( i ) b l l = 0 .infin.
a l ( i ) b - l - 1 ( 3.1 ) ##EQU00058##
is the radical inverse in an integer base b. The digit a.sub.j(i)
is the j-th digit of the index i represented in base b. The
Hammersley point set is obtained by choosing b.sub.c as the c-th
prime number. The uniformity of these points has been improved by
applying permutations to the a.sub.j(i) before computing the
inverse. The permutation
.pi..sub.b(a.sub.j(i))=a.sub.j(i)+j mod b
has been used, and other permutations have been developed,
generalizing and improving on these results. Choosing all b.sub.c=b
along with an appropriate set of mappings applied to the digits
a.sub.j(i) yields the construction and theory of (t, m, s)-nets.
There has been a lot of research as to how to efficiently compute
radical inverses. One method is to tabulate the sum of the least
significant T digits and to reuse them while generating the
points
j = 0 .infin. .pi. b ( a j ( i ) ) b - j - 1 = j = T .infin. .pi. b
( a j ( i ) ) b - j - 1 only every b T - th time + j = 0 T - 1 .pi.
b ( a j ( i ) ) b - j - 1 Table of size b T ##EQU00059##
This method has been developed in the context of scrambled radical
inversion. Rather than using Gray-codes, this method generates the
points in their natural order at comparable speed.
[0180] (2) Rank-1 Lattice Points
x i = i n ( 1 , g 1 , , g s - 1 ) mod 1 ##EQU00060##
are faster to generate than radical inversion based points. Their
quality depends on the integer generator vector
(1,g.sub.1, . . . g.sub.s-1).di-elect cons.
However, the construction of good generator vectors is not obvious.
In order to reduce the search space, the generator vectors have
been determined by only one parameter .alpha. with g.sub.i=a.sup.i.
Higher rank lattices cm be constructed by linear combinations of
rank-1 lattices.
[0181] Both principles can be generalized to yield sequences of
points, which allow for adaptive sampling without discarding
previously taken samples, however, at the price of a slight loss of
uniformity: the Halton sequence and its variations corresponding to
the Hammersley points, (t, s)-sequences containing (t, m, s)-nets,
and extensible lattice roles containing lattices.
[0182] The above constructions yield rational numbers in the unit
interval. It is especially interesting to use the base b=2 and
n=2.sup.m points, because then the points can be represented
exactly in the actual machine numbers .OR right. as defined by the
ANSI/IEEE 754-1985 standard for binary floating point
arithmetic.
[0183] The different constructions of the previous section in fact
have one common feature: They induce uniform partitions of the unit
cube. This kind of uniformity has been characterized as
follows:
TABLE-US-00001 Definition 1. Let (X, B, .mu.) be an arbitrary
probability space and let M be a nonempty subset of B. A point set
P.sub.n of n elements of X is called (M, .mu.)-uniform if i = 0 n -
1 X M ( x i ) = .mu. ( M ) n for all M .di-elect cons. M ,
##EQU00061## where XM(x.sub.i) = 1 if x.sub.i .di-elect cons. M,
zero otherwise.
[0184] Examples of (M, .mu.)-uniform point sets are samples from
the Cartesian product midpoint rule and radical inversion based
points, hi addition, rank-1 lattices are also (M, .mu.)-uniform.
The Voronoi-diagram of a lattice partitions the unit cube into n
sets of identical shape and volume
1 n . ##EQU00062##
This underlines that for (M, .mu.)-uniformity all .mu.(M) must have
the same denominator n.
[0185] The function classes of computer graphics imply to use fee
probability space
([0,1).sup.g,B,.lamda..sub.s)
with the Borel-sets B and the s-dimensional Lebesgue-measure
.lamda..sub.s.
[0186] A sequence of point sets is uniformly distributed if and
only if its discrepancy vanishes in the limit. The deterministic
constructions sketched in previous section can obtain so-called low
discrepancy, which vanishes with roughly speaking 1/n, while
independent random points only can obtain roughly
1 n ##EQU00063##
and points from the Cartesian product midpoint rule even only
acquire
1 n . ##EQU00064##
[0187] There are some facts about discrepancy that make it
problematic. Discrepancy is an anisotropic measure, because its
concept is based on axis-aligned boxes. Consequently, discrepancy
is influenced by rotating point sets. While samples from the
Cartesian product midpoint rule result in bad discrepancy, lattice
points from the Fibonacci lattices have low discrepancy, although
some of them are just rotated rectangular grids. Discrepancy is not
even shift-invariant since shifting a point set on the unit torus
also changes discrepancy.
[0188] Definition 1, above, supports partitions which are not
axis-aligned, as for example fee Voronoi-diagram of a rank-1
lattice. Maximum uniformity in this sense can be obtained by
selecting the points such that the regions of the Voronoi-diagram
approximate spheres as much as possible, i.e., by maximizing the
mutual minimum distance
d min ( P n ) := min 0 .ltoreq. i < n min i < j < n x j -
x i T ##EQU00065##
among all sample points in P.sub.n.parallel..cndot..parallel..sub.T
is used to denote the Euclidean distance on the unit torus. The
minimum distance measure is isotropic and shift-invariant thus
overcoming these disadvantages of discrepancy.
[0189] FIGS. 11A-H show a series of plots 260-330 illustrating
classical quasi-Monte Carlo points for n= 16 (top row) and n= 64
(bottom row) along with their mutual minimum distance d.sub.min.
The rank-1 lattice has been selected such that its minimum distance
is maximal. It is interesting to observe that fee constructions
with better discrepancy have larger minimum distance, too, as can
be seen for the Hammersley points, the Sobol sequence, and the
Larcher-Pillichshammer points. It also can be observed that the
minimum distance of the Halton sequence, with permutations, is
maximized as compared to the original Halton sequence.
[0190] The rank-1 lattices in FIGS. 11D and 11H are Korobov
lattices, where the parameter a has been chosen to maximize the
minimum distance. The rank-1 lattice at n=16 points in FIG. 11D in
fact can be generated as a (t, 4, 2)-net in base b=2. This is not
possible for the quality parameter t=0, because at least two points
lie on one line parallel to fee x-axis. In this case the best
quality parameter prevents the points from reaching maximum minimum
distance.
[0191] Similarly, postulating gcd(g.sub.i, n)=1 restricts the
search space for generator vectors of lattices in such a way that
the lattices with maximum minimum distance cannot be found. Forcing
Latin hypercube properties by t=0 or gcd(g.sub.i, n)=1,
respectively, may be useful in some situations, however, prevents
the sample points to optimally cover the unit torus in the minimum
distance sense.
[0192] Maximizing the minimum distance for generating highly
uniform point sets is a principle of nature. For example the
distribution of the receptors in the retina is grown that way.
Algorithmically, this principle is known as "Lloyd's relaxation
scheme," according to which samples with identical charges are
placed on fee unit torus and then the points are allowed to repel
each other until they reach some equilibrium. From a mathematical
point of view, the points are moved to the center of gravity of
their Voronoi cells during relaxation. The convergence of this
scheme has been improved quadratically. Note that rank-1 lattices
are invariant under this kind of relaxation scheme.
[0193] Point sets selected by maximizing the mutual Euclidean
distance on the unit torus induced by
.parallel..cndot..parallel..sub.T have the advantage that they
seamlessly can be tiled in order to fill s-dimensional space. This
property is intrinsic, to lattices. However, it is not true for
radical inversion based points in general. While the points by
Larcher and Pillichshammer tile seamlessly, the Hammersley points
do not. The norm .parallel..cndot..parallel..sub.T in fact should
be a weighted norm, which includes the size of the integration
domain. FIGS. 12A-C shows a series of drawings 340-360,
illustrating selection of lattices by maximum minimum distance in
fee unit cube (FIG. 12A); in the unit cube scaled to the
integration domain (FIG. 12B); and (in the integration domain (FIG.
12C). Considering the scale of the integration domain yields more
uniform points. In particular, it will be see that maximizing the
minimum distance on the unit cube does not imply nice uniformity
when scaling the domain of integration.
[0194] It is common knowledge that quasi-Monte Carlo integration
outperforms Monte Carlo integration in many applications. However,
the classical error bound theorems often do not fit the function
classes of the application.
[0195] The Koksma-Hlawka inequality deterministically bounds the
integration error by the product of the discrepancy of the sample
points used to determine the sample mean and the variation of the
integrand in the sense of Hardy and Krause. The variation can be
considered as the remainder of trying to obtain discrepancy as a
factor in the error bound.
[0196] The class of functions of bounded variation is impractical
in certain situations. For example, discontinuities that are not
aligned with the coordinate axes already yield infinite variation.
The error bound thus becomes useless in already simple settings of
computer graphics such as an edge
f ( x , y ) = { 1 y > x 0 else ( 3.2 ) ##EQU00066##
in the unit square, for which the variation is unbounded. Using
other kinds, such as for example isotropic discrepancy, it is
possible to find an error bound that works for this case. However,
it becomes far too pessimistic in higher dimensions.
[0197] Similar to the Koskma-Hlawka inequality there exist
deterministic error bounds for the integration by lattices. The
used function class requires periodic functions and imposes certain
constraints on the Fourier coefficients of the integrand, which do
not apply for discontinuous functions as used in computer
graphics.
[0198] In a vast number of publications experiments give numerical
evidence that quasi-Monte Carlo methods outperform Monte Carlo
methods in practice, however, in the majority of the cases
classical theorems cannot explain, the observed results. The main
reason is that general discontinuities cannot be accounted for by
the classical error bounds.
[0199] As stated above, image synthesis is an Integro-approximation
problem in L.sub.b.sup.2 and quasi-Monte Carlo
integro-approximation in fact successfully has been used in
computer graphics. Therefore, the following theorem may be
generalized in the sense of parametric integration:
TABLE-US-00002 Theorem 1. Let (X, B, .mu.) be an arbitrary
probability space and let M = {M.sub.1, . . . ,M.sub.k} be a
partition of X with M.sub.j .di-elect cons. B for 1 .ltoreq. j
.ltoreq. k. Then for any (M, .mu.)-uniform point set P = {x.sub.1,
. . . ,x.sub.n} and any bounded function f, which restricted to X
is .mu.-integrable, we have 1 n i = 0 n - 1 f ( x i , y ) - .intg.
x f ( x , y ) .mu. ( x ) .ltoreq. ##EQU00067## j = 1 k .mu. ( M j )
sup x .di-elect cons. M j f ( x , y ) - inf x .di-elect cons. M j f
( x , y ) ##EQU00068## for any suitable norm .parallel.
.parallel..
[0200] This theorem may be proved as follows: For all y.di-elect
cons.Y, consider an arbitrary stratum M.sub.j.di-elect cons.M.
Then,
i = 0 n - 1 .chi. M j ( x i ) inf x .di-elect cons. M j f ( x , y )
.ltoreq. i = 0 x i .di-elect cons. M j n - 1 f ( x i , y ) .ltoreq.
i = 0 n - 1 .chi. M j ( x i ) sup x .di-elect cons. M j f ( x , y )
##EQU00069##
which implies
.mu. ( M j ) inf x .di-elect cons. M j f ( x , y ) .ltoreq. 1 n i =
0 x i .di-elect cons. M j n - 1 f ( x , y ) .ltoreq. .mu. ( M j )
sup x .di-elect cons. M j f ( x , y ) ##EQU00070##
because P is an (M, .mu.)-uniform point set. Similarly,
.mu. ( M j ) inf x .di-elect cons. M j f ( x , y ) .ltoreq. .intg.
M j f ( x , y ) .mu. ( x ) .ltoreq. .mu. ( M j ) sup x .di-elect
cons. M j f ( x , y ) ##EQU00071##
From the latter two sets of inequalities it follows that
1 n i = 0 x i .di-elect cons. M j n - 1 f ( x i , y ) - .intg. M j
f ( x , y ) .mu. ( x ) .ltoreq. .mu. ( M j ) ( sup x .di-elect
cons. M j f ( x , y ) - inf x .di-elect cons. M j f ( x , y ) )
##EQU00072##
Since M is a partition of X,
1 n i = 0 n - 1 f ( x i , y ) - .intg. X f ( x , y ) .mu. ( x ) = 1
n j = 1 n - 1 i = 0 x i .di-elect cons. M j n - 1 f ( x i , y ) - j
= 1 k .intg. M j f ( x , y ) .mu. ( x ) = j = 1 k ( 1 n i = 0 x i
.di-elect cons. M j n - 1 f ( x i , y ) - .intg. M j f ( x , y )
.mu. ( x ) ) ##EQU00073##
[0201] Using the previous inequality and applying the norm
.parallel..cndot..parallel. to both sides of the resulting
inequality yields the desired bound.
[0202] By the omission of y, the norm reduces to the absolute value
and it remains the original theorem and proof of the following
Theorem 2:
TABLE-US-00003 Theorem 2. Let (X, B, .mu.) be an arbitrary
probability space and let M = {M.sub.1, . . . ,M.sub.k} be a
partition of X with M.sub.j .di-elect cons. B for 1 .ltoreq. j
.ltoreq. k. Then for any (M, .mu.)-uniform point set P = {x.sub.0,
. . . , x.sub.n-1} and any bounded .mu.-integrable function f on X
we have .intg. X f ( x ) .mu. ( x ) - 1 n i = 0 n - 1 f ( x i )
.ltoreq. j = 1 k .mu. ( M j ) ( sup x .di-elect cons. M j f ( x ) -
inf x .di-elect cons. M j f ( x ) ) ##EQU00074##
[0203] By using the concept of (M, .mu.)-uniform point sets instead
of discrepancy, proofs become simpler and results are more general,
compared with earlier approaches. With (X, B, .mu.)=([0, 1).sup.s,
B, .lamda..sub.s) both theorems are applicable in the setting of
computer graphics.
[0204] For the example, the deterministic error O(n.sup.-1/2) bound
can be obtained by selecting an (M, .mu.)-uniform point set with
k=n. The difference of the supremum and the infimum can only be one
in the O(n.sup.-1/2) sets of the partition, which are crossed by
the discontinuity, otherwise it must be zero. With .mu.(Mj)=1/n,
the bound is obtained. It will be seen that this argument does not
use probabilistic arguments for a deterministic algorithm.
[0205] Quasi-Monte Carlo methods are biased, because they are
deterministic, but consistent, because they asymptotically converge
to the right solution. Randomizing these algorithms allows for
unbiased estimators and for unbiased error estimators.
[0206] It is useful to consider the method of dependent tests
.intg. [ 0 , 1 ) s f ( x , y ) x = .intg. [ 0 , 1 ) s i = 0 n - 1 w
i ( x , y ) f ( R i ( x ) , y ) x .apprxeq. i = 0 n - 1 w i (
.omega. , y ) f ( R i ( .omega. ) , y ) ( 3.3 ) ##EQU00075##
by first applying an equivalence transformation, which does not
change the result, and then using one random sample .omega. in
order to obtain an unbiased estimate.
[0207] This formulation is able to cover a wide range of techniques
that increase the efficiency of the method of dependent tests. For
example, random scrambling of arbitrary points, random translation
on the unit torus, random padding of arbitrary point sets,
stratification and stratification induced by rank-1 lattices,
trajectory splitting, and many more techniques easily can be
formulated by a tuple of replications R.sub.i with associated
weights w.sub.i.
[0208] Omitting y, the special case with equal weights
w i ( x ) = 1 n ##EQU00076##
where for fixed .omega. the set of points
(R.sub.i(.omega.)).sub.i=0.sup.n-1
R.sub.i(.omega.))n-1 is of low discrepancy, is defined to be
randomized quasi-Monte Carlo integration.
[0209] Repeating independent realizations allows one to estimate
the error of the approximation in an unbiased way. However, some
convergence is sacrificed, since independent random samples cannot
be as uniformly distributed as correlated samples can be. This is
especially noticeable in the setting of computer graphics, where
small numbers of samples are used.
[0210] Anti-aliasing is a central problem in computer graphics.
FIG. 13 is a computer-generated image 370 of an infinite plane with
a checkerboard texture that illustrates various difficulties. While
in the front of the checkerboard, the fields clearly can be
distinguished, extraneous patterns appear towards the horizon.
While it is simple to compute the color of a pixel as an average as
long as the checker board cells are clearly distinguishable, this
is no longer possible at fee horizon, where through one pixel
infinitely many cells can be seen. By common sense, it would be
expected that, the pixels at the horizon would be gray, i.e., the
average of black and white. However, surprisingly, zooming into a
pixel reveals that the areas of black and white tiles are not equal
in general. This means that no matter what quadrature is used, the
horizon will not appear gray, but somehow patterned.
[0211] The lens of the human eye is not perfectly transparent and
thus slightly blurs the light before it reaches fee retina. The
amount of blur perfectly fits the resolution of fee receptors in
the retina. A similar trick in computer graphics is to blur the
texture before integration, where the strength of the blur depends
on the distance from the eye. However, tins blurring technique does
not help, if fee mathematical problems are not caused by textures.
Another compromise is to filler the resulting image. However, a
filtering technique causes not only aliases but also previously
sharp details to become blurred.
[0212] Aliasing can only be hidden by random sampling. Taking one
independent random sample inside each pixel, the horizon will
appear as uncorrected black and white pixels. The structured
aliases thus are mapped to noise, which is less disturbing, to the
eye. However, taking more random samples per pixel, the quadrature
will eventually converge and aliases will appear instead of the
noise.
[0213] Assuming that situations like the ones mentioned before can
be managed by suitable filtering, there are now discussed
alternative sampling patterns of computer graphics. These sampling
patterns are illustrated in FIGS. 14A-C. FIG. 14A shows stratified
sampling 380; FIG. 14B shows Latin hypercube sampling 390; and FIG.
14C shows blue noise sampling 400. FIGS. 15A-E show a series of
drawings 410-450, illustrating sampling using quasi-Monte Carlo
points.
[0214] Stratified sampling, shown in (a), is at least as good as
random sampling. However, stratified sampling suffers from the
curse of dimension, since each axis of the integration domain has
to be divided at least once.
[0215] Latin hypercube sampling, shown in (b), can never be much
worse than random sampling and is available for any dimension and
number of samples. The average observed performance is good,
although it cannot be guaranteed.
[0216] Poisson disk sampling, shown in (c), simulates the
distribution of receptors in the human eye. The mutual minimum
distance of the points is maximized, which results in a reduced
noise level in the rendered image. Although restricted by a
guaranteed minimum distance, i.e., an empty disk around each
sample, the samples points are randomly placed. Thus representable
details remain sharp in the final image and aliases are efficiently
mapped to noise. However, Poisson disk sampling patterns are
typically expensive to generate.
[0217] The properties of the sampling patterns seem disjointed,
however, there exist quasi-Monte Carlo points, which can be
generated efficiently, and which unite the above properties.
[0218] There is now discussed an anti-aliasing technique, using (0,
2m, 2)-nets in base b=2.
[0219] The use of Hammersley points in the accumulation buffer
algorithm has been investigated. Despite considerable improvements,
each pixel had to use the same samples, which required higher
sampling rates in order to avoid aliases. One solution to that
problem is to exploit the structure of (0, 2m, 2)-nets in base
b=2.
[0220] FIGS. 15A-E illustrate that (0, 2m, 2)-nets in base b=2
unite the properties of classical sampling patterns in computer
graphics. The first two dimensions
x i = ( i ( 2 m ) 2 , .PHI. 2 ( i ) ) ##EQU00077##
of the Hammersley points are an example of such a point set, which
is stratified, a Latin hypercube sample, and has a guaranteed
minimum distance of at least
1 ( 2 m ) 2 . ##EQU00078##
[0221] The method of dependent tests is realized by tiling the
image plane with copies of the (0, 2m, 2)-net. FIG. 16 shows a plot
460, illustrating how the samples in a pixel are determined by
tiled instances of a Hammersley point set. The solid lines indicate
the unit square including one set of 16 Hammersley points and the
dashed lines indicate screen pixel boundaries. Because of a lack of
space, the illustration uses only 4 samples per pixel.
[0222] In order to reduce aliasing artifacts, neighboring pixels
should have different samples. This is achieved by letting one
Hammersley point set cover multiple pixels. The improvement in
convergence is visible in the same figure, where we compare
stratified random sampling and anti-aliasing by using the
Hammersley point set. The improvements directly transfer to other
rendering algorithms, such as for example the REYES architecture as
used in PIXAR's RenderMan software.
[0223] Selecting rank-1 lattices by maximum minimum distance takes
the principle of Lloyd relaxation and Poisson disk sampling to the
limit. Since rank-1 lattices are available for any number n of
points, the sampling rate can be chosen freely. A factorization of
n as required for axis-aligned stratified sampling is not needed.
In order to attenuate aliasing in each pixel a different random
shift is added to the lattice points resulting in an unbiased
estimate. The images obtained that way display minimal noise and
aliasing. The samples of a randomly shifted lattice can be
generated fast and the observed convergence is even faster as
compared to the radical inversion based methods from the previous
section. The method can be derandomized by extracting the shift
from, e.g., a Hammersley point set using the stratification
properties pointed out in the previous section.
[0224] The principles of the previous sections can be extended to
approximate the full integro-approximation problem. Using the
Neumann series the computation of L is reduced to a sum of
integrals as sketched in fee introduction. Extensive investigations
have been carried out on path tracing algorithms in connection with
quasi-Monte Carlo methods and randomized versions of the latter
methods. All the experiments resulted in improvements when using
quasi-Monte Carlo methods as compared to classical algorithms of
computer graphics.
[0225] As mentioned above, randomizing quasi-Monte Carlo methods
allows for unbiased estimators and unbiased error estimates. The
latter does not appear to be particularly interesting in computer
graphics, as better adaptive methods already exist, as discussed
below. Furthermore fee resulting images can be generated in the
same time and quality no matter whether a quasi-Monte Carlo or its
randomized counterpart is used.
[0226] However, it is interesting to look at the effect of
randomization schemes on the minimum distance. While randomizing
point sets by random shifts does not change their maximum minimum
distance on the torus, random scrambling does.
[0227] Applying random scrambling to the classical low discrepancy
points sets it can be observed that often the uniformity is
improved and rarely decreased. This is true for both discrepancy
and minimum distance. Experiments indicate that random scrambling
only marginally changes the uniformity of the scrambled Halton
sequence, indicating that deterministic permutations already are a
very good choice.
[0228] Permutations for the Halton sequence are realizations of
random scrambling and can be considered as deterministic scrambling
long before random scrambling was introduced. Although the
permutations have been introduced to improve on discrepancy, they
also increase the maximum minimum distance of the Halton points.
Implementing the scrambled Halton sequence is much simpler than
random scrambling and avoids the risk that uniformity if influenced
negatively.
[0229] In computer graphics, the difficulty of the
integro-approximation problem differs from pixel to pixel and
adaptive methods pay off. A possible criterion for refinement is to
compare the image gradient
.parallel..gradient.I(k,l).parallel..sub.2>Tn.sub.k,l.sup.-.alpha.
(3.4)
against a threshold T. By the exponent .alpha..di-elect cons.[0.5,
1.5] we can adapt to the speed of convergence and n.sub.k,l is the
number of samples used for pixel (k, l). Alternatively refinement
can be indicated by the routines that evaluate the radiance L,
since these routines can access more information of the scene to be
rendered.
[0230] In the following discussion, there is described a technique
for adaptive anti-aliasing by elements of the Halton sequence. The
method is superior to random sampling, since the points of the
Halton sequence are more uniform and spoken visually exactly fail
into the largest unsampled gaps of the sampling domain resulting in
a smooth convergence.
[0231] The radical inverse in Equation (3.1) mirrors the
representation of the index i by the digits
.alpha..sub.i(i).di-elect cons.{0, . . . , b-1} in the integer base
b.di-elect cons.N at the decimal point. In base b=2 this means
that
.PHI. 2 ( i ) .di-elect cons. { [ 0 , 1 / 2 ) for i even , [ 1 / 2
, 1 ) else . ( 3.5 ) ##EQU00079##
[0232] This observation has been generalized and named periodicity.
In fact, however, this property relates much more to stratification
as formalized by the definition of (0, 1)-sequences. Therefore, a
different derivation is presented that stresses the actual
stratification property. The index i is chosen as
i=jb.sup.n+k for k.di-elect cons.{0, . . . b.sup.n-1}
and inserted into (1), yielding
.PHI. b ( i ) = .PHI. b ( j b n + k ) = l = 0 .infin. a l ( j b n +
k ) b - l - 1 = l = 0 .infin. a l ( j b n ) b - l - 1 + l = 0
.infin. a l ( k ) b - l - 1 = l = 0 .infin. a l ( j ) b - n - l - 1
+ .PHI. b ( k ) = b - n .PHI. b ( j ) .di-elect cons. [ 0 , b - n )
+ .PHI. b ( k ) ( 3.6 ) ##EQU00080##
by exploiting the additivity of the radical inverses that belong to
the class of (0, 1)-sequences. The first term of the result depends
on/and obviously is bounded by b.sup.-n, while the second term is a
constant offset by the radical inverse of k. Since
k.di-elect cons.{0, . . . , b.sup.n-1}
it follows that
.PHI..sub.k.di-elect cons.{0,b.sup.-n,2b.sup.-n, . . . ,
(b.sup.n-1)b.sup.-n}
Fixing the n first digits by k then for j.di-elect cons.N .sub.0
yields radical inverses only in the interval
[.PHI..sub.b(k),.PHI..sub.b(k)+b.sup.-n)
[0233] The one-dimensional observations from She previous section
generalize to higher dimensions. For the Halton sequence
x.sub.i=(.PHI..sub.b.sub.g(i), . . . ,
.PHI..sub.b.sub.g(i)).di-elect cons.[0,1).sup.g
where for the c-th component b.sub.c is the c-th prime number, this
is seen by choosing the index
i = j d = 1 s b d n d + k with 0 .ltoreq. k < d = 1 s b d n d (
3.7 ) ##EQU00081##
yielding
.PHI. b c ( i ) = .PHI. b c ( j d = 1 s b d n d + k ) = b c - n c
.PHI. b c ( j d = 1 d .noteq. c s b d n d ) + .PHI. b c ( k )
##EQU00082##
which is analogous to Equation (3.6) and consequently
x i .di-elect cons. c = 1 s [ .PHI. b c ( k ) , .PHI. b c ( k ) + b
c - n c ) ##EQU00083##
for fixed k and j.di-elect cons..sub.0 with the choice of the
index/according to Equation (3.7). The
.PI..sub.d=1.sup.gb.sub.d.sup.n.sup.d
disjoint strata selected by k are disjoint and form a partition of
the unit cube 10, [0, 1).sup.s. However the scheme suffers from the
curse of dimension, since stratifying all s-dimensions results in
an exponential growth of the number of strata. The increment
.PI..sub.d=1.sup.sb.sub.d.sup.n.sup.d
grows even foster than 2.sup.s since except the first, all
b.sub.d>2.
[0234] A tensor product approach has been compared with an approach
using the Halton sequence. An illustration has been constructed of
the stratification of path space for photons that are emitted on
the light source and traced for 2 reflections. The problem requires
samples from the 8-dimensional unit cube. The sampling domain was
stratified into 128.sup.8 strata of identical measure. Then one
stratum was selected and 8 random samples have been drawn from if
to determine the photon trajectories. Stratification by the Halton
sequence was been used to determine the 8 paths. In spite of an
enormously large increment that hardly fits into the integer
representation of a computer, the trajectories start to diverge
after the 2nd reflection.
[0235] For a stratification as fine as the tensor product approach
large increments
.PI..sub.d=1.sup.sb.sub.d.sup.n.sup.d
are required that hardly fit the integer representation on of a
computer. In addition, the assumption that close points in the unit
cube result in close paths in path space is not valid for more
complex scenes. Depending on what part of the geometry is hit the
photons can be scattered in completely different parts of the
scene, although their generating points bad been close in the unit
cube. The longer the trajectories the more diverging they will be.
The above observations easily can be transferred to (t,
s)-sequences in base b.
[0236] While stratification by the Halton sequence is not useful in
high dimensions, it can be very useful in small dimensions as, for
example, in pixel anti-aliasing. The properties in two dimensions
are illustrated in FIGS. 17A and 17B, which are plots 470 and 480
illustrating how the samples from the Halton sequence in the unit
square were scaled to fit the pixel raster. The plots show,
respectively, the first two components x.sub.i=(.PHI..sub.2(i),
.PHI..sub.3(i)) of the Halton sequence for
0.ltoreq.i<2.sup.33.sup.3=216. The solid points have the indices
i.ident.i.sub.k(j)=1.sup.13.sup.1j+k selected by k=1. The stratum
with the emphasized points contains all indices i.ident.i.sub.k(j)
with k=1. In order to match the square pixels the coordinates are
scaled, i.e.,
x.sub.i.fwdarw.x'.sub.i=(2.sup.1.PHI..sub.2(i),3.sup.1.PHI..sub.3(i))
[0237] In general the first component x.sub.i.sup.(1) is scaled by
b.sub.1.sup.n.sup.1 and the second x.sub.i.sup.(2) by
b.sub.2.sup.n.sup.2. Thus a
b.sub.1.sup.n.sup.1.times.b.sub.2.sup.n.sup.2 stratified sample
pattern is obtained that can be periodically tiled over the image
plane, as described above. Identifying each stratum with a pixel,
the identification k easily is determined (for example by a table
lookup) by the pixel coordinates and a Halton sequence restricted
to that pixel is obtained from
i.ident.i.sub.k(j)=jb.sub.1.sup.n.sup.1b.sub.2.sup.n.sup.2+k for
j.di-elect cons..sub.0
[0238] Exemplary images have been computed using a path tracer, as
described above, with the scrambled Halton sequence. Refinement was
triggered by the gradient (4). Note that the scrambling does not
change .PHI..sub.2 and .PHI..sub.3. Consequently, the above
algorithm can be applied directly and benefits from the improved
uniformity of the scrambled sequence.
[0239] Trajectory splitting can increase efficiency in rendering
algorithms. A typical example is volume rendering. While tracing
one ray through a pixel it is useful to send multiple rays to the
light sources along that ray. It has been shown that taking
equidistant samples on fee ray that all have been randomly shifted
by fee same amount is much more efficient than the original method
that used jittered sampling.
[0240] A general approach to trajectory splitting is to restrict
the replications in Equation (3.3) to some dimensions of the
integrand. For quasi-Monte Carlo methods, this approach has been
used in an implementation, according to which a strictly
deterministic version of distribution ray tracing was developed. A
systematic approach has been taken, according to which
randomization techniques from fee field of randomized quasi-Monte
Carlo methods have been parameterized. Instead of using random
parameters, deterministic quasi-Monte Carlo points have been
applied. Seen from a practical point of view, trajectory splitting
can be considered as low-pass filtering of the integrand with
respect to the splitting dimensions.
[0241] The most powerful method is to split trajectories using
domain stratification induced by rank-1 lattices. For the
interesting s-dimensions of the problem domain, a rank-1 lattice is
selected. The matrix B contains the vectors spanning fee unit cell
as identified by the Voronoi-diagram of the rank-1 lattice.
Then
R i : [ 0 , 1 ) s .fwdarw. A i ##EQU00084## x ( i n ( 1 , g 1 , , g
s - 1 ) + Bx ) mod [ 0 , 1 ) s ##EQU00084.2##
maps points from fee unit cube to the i-th stratum A, of the rank-1
lattice, as depicted in FIG. 18B, discussed below. This scheme can
be applied recursively yielding recursive Korobov filters.
[0242] For the special case of the Fibonacci lattice at n=5 points
the recursive procedure has been used for adaptive sampling in
computer graphics. Starting with the lattice Z.sup.2, the next
refinement level is found by rotating Z.sup.2 by arctan(1/2) and
scaling it by 1/ {square root over (5)}, as indicated in FIG. 18C,
discussed below. The resulting lattice again is a rectangular
lattice and the procedure recursively can be continued. Thus, the
construction is completely unrelated to rank-1 lattices.
[0243] FIGS. 18A-C show a series of plots 490-510 illustrating
replications by rank-1 lattices. The Voronoi-diagram of a rank-1
lattice induces a stratification, shown in FIG. 18A. All cells
A.sub.g are of identical measure and in fact rank-1 lattices are
(M, .mu.)-uniform. A cell A.sub.g is anchored at the i-th lattice
point x.sub.i and is spanned by the basis vectors (b.sub.1,
b.sub.2). This can be used for recursive Korobov-filters, shown in
FIG. 18B, where the points inside a lattice cell are determined by
another set of lattice points transformed into that lattice cell.
In computer graphics one special case (c) of this principle has
been named {square root over (5)} sampling, shown in FIG. 18C,
because the length of the dashed lines is 1/ {square root over
(5)}. It is in fact a recursive Korobov filter with points from the
Fibonacci lattice at n=5 points.
[0244] Good results have been obtained from a distribution ray
tracer that used randomly shifted rank-1 lattices with maximized
minimum distance. Trajectory splitting was realized using rank-1
lattices with maximized minimum distance, too.
[0245] One random vector was used per pixel to shift the lattice
points in order to obtain an unbiased estimator and to decorrelate
neighboring pixels. The resulting images exposed minimal noise and
while aliasing artifacts are pushed to noise. Compared to previous
sampling methods the convergence was superior, due to the high
uniformity of the lattice points.
[0246] As lattice points are maximally correlated, this is a good
example for correlated sampling in computer graphics. In this
context; quasi-Monte Carlo integro-approximation by lattice points
can be considered as Korobov filtering.
[0247] While applications of quasi-Monte Carlo integration in
finance attracted a lot of attention instantly, developments in
computer graphics were not that spectacular. Today, however, about
half of the rendered images in movie industry are synthesized using
strictly deterministic quasi-Monte Carlo integro-approximation. In
2003, these techniques were awarded a Technical Achievement Award
(Oscar) by the American Academy of Motion Picture Arts and
Sciences. In contrast to academia, graphics hardware and software
industry early recognized the benefits of quasi-Monte Carlo
methods.
[0248] Deterministic quasi-Monte Carlo methods have the advantage
that they can be parallelized without having to consider
correlation as encountered when using pseudo-random number
generators. By their deterministic nature the results are exactly
reproducible even in a parallel computing environment.
[0249] Compared to classical algorithms of computer graphics the
algorithms are smaller and more efficient, since high uniformity is
intrinsic to the sample points. A good example is trajectory
splitting by rank-1 lattice that have maximized minimum
distance.
[0250] In computer graphics it is known that maximizing the minimum
distance of point sets increases convergence speed. However,
algorithms to create such points, such as Lloyd's relaxation
method, are expensive. With quasi-Monte Carlo methods selected by
maximized minimum distance, efficient algorithms are available and
savings up to 30% of the computation time for images of the same
quality as compared to random sampling methods can be observed.
[0251] In the selling of computer graphics quasi-Monte Carlo
methods benefit from the piecewise continuity of the integrands in
L.sub.b.sup.2. Around the lines of discontinuity the methods are
observed to perform no worse than random sampling, while in the
regions of continuity the better uniformity guarantees for faster
convergence. The observed convergence rate is between O (n.sup.-1)
and O(n.sup.-1/2). It depends on the ratio of the number of sets in
the partition induced by (M, .mu.)-uniform points and the number of
these sets containing discontinuities. Since with increasing number
of dimensions the integrands tend to contain more discontinuities
the largest improvements are observed for smaller dimensions.
[0252] Since photorealistic, image generation comprises fee
simulation of light transport by computing functionals of the
solution of a Fredholm integral equation of the second kind, the
quasi-Monte Carlo methods developed for computer graphics apply to
other problems of transport theory as well.
[0253] The concept of maximized minimum distance as used in
computer graphics nicely fits the concept of (M, .mu.)-uniformity
as used in quasi-Monte Carlo theory. Rank-1 lattices selected by
maximized minimum distance ideally fit both requirements and yield
superior results in computer graphics.
IV. General Methods
[0254] FIGS. 19-22 are a series of flowcharts illustrating general
methods according to further aspects of the present invention.
[0255] FIG. 19 is a flowchart of a computer-implemented method 600
for generating a pixel value for a pixel in an image displayable
via a display device, the pixel value being representative of a
point in a scene. The method includes the following steps:
[0256] Step 601; Generating a set of sample points, at least one
sample point being generated using at least one sample, the at
least one sample comprising at least one element of a sequence, and
wherein the set of sample points comprises quasi-Monte Carlo
points.
[0257] Step 602: Evaluating a selected function at one of the
sample points to generate a value, the generated value
corresponding to the pixel value, the pixel value being usable to
generate a display-controlling electronic output.
[0258] FIG. 20 shows a flowchart of a computer-implemented method
620 for generating a pixel value for a pixel in an image display
able via a display device, the pixel value being representative of
a point in a scene. The method comprises the following steps:
[0259] Step 621: Generating a set of sample points, at least one
sample point being generated using at least one sample, the at
least one sample comprising at least one element of a
low-discrepancy sequence, and wherein the generating includes using
an adaptive, interleaved sampling scheme based on a
deterministically scrambled Halton sequence to yield a
deterministic, low-discrepancy set of sample points.
[0260] Step 622: Evaluating a selected function at one of the
sample points to generate a value, the generated value
corresponding to the pixel value, the pixel value being usable to
generate a display-controlling electronic output.
[0261] FIG. 21 shows a flowchart of a computer-implemented method
640 for generating a pixel value for a pixel in an image
displayable via a display device, the pixel value being
representative of a point in a scene. The method comprises the
following steps:
[0262] Step 641: Generating a set of sample points, at least one
sample point being generated using at least one sample, the at
least one sample comprising at least one element of a sequence,
wherein the set of sample points comprises quasi-Monte Carlo
points, and wherein the generating includes adaptively sampling by
using radical inversion-based points.
[0263] Step 642: Evaluating a selected function at one of the
sample points to generate a value, the generated value
corresponding to the pixel value, the pixel value being usable to
generate a display-controlling electronic output.
[0264] FIG. 22 shows a flowchart of a computer-implemented method
660 for generating a pixel value for a pixel, in an image
displayable via a display device, the pixel value being
representative of a point in a scene. The method comprises the
following steps:
[0265] Step 661: Generating a set of sample points, at least one
sample point being generated using at least one sample, the at
least, one sample comprising at least one element of a sequence,
wherein the generating includes sampling by using rank-1 lattice
points.
[0266] Step 662: Evaluating a selected function at one of the
sample points to generate a value, the generated valise
corresponding to the pixel value, the pixel value being usable to
generate a display-controlling electronic output.
[0267] The foregoing description provides detail of various
embodiments, practices and examples of the invention. It will be
understood that various additions, variations and modifications may
be made to the invention, within the spirit and scope of the
present invention, the scope of which is limited solely by the
appended claims.
* * * * *