Feature Article

Carole Blanc and Christophe Schlick

University of Bordeaux

ne argument often given to explain the Conics, circles, spheres, and splines
popularity of NURBS (nonuniform ratio- We briefly review here how conic sections (and relat-
nal B-spline) is that it permits the ed curves and surfaces) can be modeled by splines.
definition of free-form curves and
Zigzag reparametrization surfaces (as do most spline models), Conics
It also provides an exact representa- In the early 1800s,Moebius stated that a conic can be
improves the tion of conic sections and thus of a considered the projection on IR2of a parabola defined in
large set of curves and surfaces used IR3.Using this result, we can show that a conic can also be
parametrization of a NURBS intensivelyincm: circulararcs, cir- expressed as a rational quadratic Bezier curve (and alter-
cles, cylinders, cones, spheres, sur- natively as a quadratic NURBS) .3,4 Such a curve is defined
curve or surface according faces of revolution, and so forth. by three points, PO,PI, Pz, and three weights, WO, w1,w2:
Nevertheless, few published works
to a given criterion, as discuss the mathematical properties
behind the representation of conics
illustrated here through by NURBS except for two recent
monographs by Piegl and Tiller' and
parametrization of a circle. by Farin.' This article does not pre-
tend to fill this theoretical lack but
rather intends to deal with the following problems: or by explicating the Bernstein polynomials Bg(t), B?(t),
and Bz(t):
All known NURBS representations of curves and sur-
faces based on conics have only a C1 continuity. kft E 10.11
Moreover, no technique exists that would eventually
allow us to find a parametrization with a higher level
of continuity. Ck)= 07
wo 1 - t P0+2w1 1-t tP1+w,
0 t 2 P2
(2)
The parametrization resulting from the NURBS rep-
resentation of conics can deviate significantly from
w,-2t w -wl
( 0 1
+ t 2 wo-2w1+w
( 1
the ideal arc length (that is, uniform) parametriza- Parameters wo,w1,and w2 are linked by the conic
tion. The only known solution to reduce this devia- invariant5 relationship wowl/wf which means that every
tion is to increase the number of control points of the set (WO, w1, w2)that provides the same value for the
spline (by using refinement algorithms, for instance), conic invariant also provides the same c ~ r v e . ~ , ~ , ~
but such a process converges only slowly to the uni- Moreover, because the roles of Po andPzare symmetric,
form parametrization. it is preferable to have wo = W Z .Finally, without loss of
generality, we can assume that wo= w2= 1and w1= w
The solution we propose here uses an original repa- (this condition is sometimes called the normalform of
rametrization process that we call zigzag reparametriza- the rational curve7),which leads to
tion, based on a specific family of rational polynomials.
This technique improves the parametrization by mini-
mizing a given error criterion (in our case, giving a high-
er order continuity as well as a more uniform Po+2w 1-t t P1+t2 Pz
parametrization). On the other hand, it raises the degree C ( t ) = (1-t) ( 1
)
1-2 4 1- w + 2t2 (1- w )
(3)
of the resulting NURBS (in our case, from two to four).

November 1996 0272-1 7-1 6/96/$5.00 0 1996 IEEE

 This expression enables a simple classificationof the
resulting curve (as shown in Figure 1):

if w < 1,then C ( t ) is an ellipse, 1 Three types

if w = 1,then C(t) is a parabola, and of conics.
if w > 1,then C(t) is a hyperbola.

Circular arcs
For geometric modeling, the circular arc is undoubt-
edly the most useful particular case of conic sections.
Such an arc results when Po, PI, PZform an isoscelestri-
angle and when w = cos@(where @ is the angle between
P& and P&).3,4,6For instance, a circular arc of length
29 starting from the trigonometric origin (see Figure 2)
is obtained with the following control points:

p2 (cos24, sin2qj
2 Circular arc.

which can be computed more efficientlyby

where T = tan@.

Circles and spheres

Equation 5 gives us a control triangle PoPlPz that only 3 Three
permits the definition of a circular arc sweeping less different
than 180 degrees (T becomes infinite for 24, = 7c). control lattices
Therefore, at least three circular arcs are needed to defining a
obtain a full circle. These arcs are pieced together (the circle.
last point of each arc becomes the first point of the next
one), giving a circle defined by n arcs and 2n+ 1control (4 (4
points. The resulting curve can be considered either a
piecewise rational Bezier or a NURBS. Arcs of different
lengths may be pieced together, but concatenating sim-
ilar arcs provides a much better parametrization. PO (1’0 ) p4
-1
-,-
-&
Several authors have proposed representations for
circular arcs sweeping over 180 degrees by using infi- Pl(l>&] (2
nite control points or negative weights.6,8Such con- P 5 d q
structions are less relevant to CAD, because they lose
the convex hull property and give bad parametrizations.
The reparametrization process we propose, on the other
;[ f ]
P2 -,-

p3 (-2,o)
Pb(L0)

hand, may eventuallybe employed to improve the qual-

itv of this Darametrization.
Among all the possible representations, the follow. where even-indexed (or, respectively, odd-indexed)
ing ones appear particularly interesting. points have their weights set to 1(respectively, 1/2).

Triangle-basedcircle. Versprille4 proposed the Square-basedcircle. The main drawback of the

first exact representation of a circle by a NURBS. In that triangle-based circle is that the size of the control poly-
model, the circle is composed of three arcs (each of gon is relatively large compared to the circle. An
length 2~c/3)and defined by six control points (one improvement proposed by Tiller’ defines the circle by
point being repeated) regularly placed on an equilater- four arcs (each of length x/2) using eight control points
al triangle (see Figure 3a): (one repeated) placed on a square (see Figure 3b):

IEEE Computer Graphics and Applications 65

Feature Article

circles (providingthe meridians). The most popular rep-

resentation, proposed by Tiller,' is based on the cir-
4 Two different cumscribing cube (defined with 26 control points, five
control lattices of them repeated). We can also imagine a tighter con-
defining a trol lattice having a kind of hexahedrical shape (defined
sphere. with 62 control points, seven of them repeated), as
shown in Figure 4.

Since the circle results from concatenation of several

circular arcs, the question of continuity at the junction
points (knots in B-spline terminology) arises naturally.
5 junction of To study this problem," let us take the case where two
two circular arcs, defined by the control trianglesP&Pz andPgP4,
arcs. respectively (see Figure 5), are connected. When the
arcs have the same length, we necessarily have wo= wz
= w4 = 1and w1 = w3 = w.Moreover, let us suppose
that the parameter range is [0,1] for the first segment
and [ 1,2] for the second one. If we use the 1-and I+
notation to express the left and right limits for the curve
at the parameter value t = 1,Equation 3 gives

C+P2

C'(7j=2w jp2 -5)

( ) =2i
c"1- o 1 i
P -Pz + 4 w 1-w PI-P,)

C(l')=Pz
where even-indexed (respectively, odd-indexed) points
have their weights set to 1(respectively, d2/2). Cf$')=2wjP,-P2)

Hexagon-basedcircle.The square-based circle is C j l +)=2 iP4 - Pz )+ 4w (1- w)(p3- P2 )

by far the most commonly used in modeling software,
mainly because of its simplicity. Nevertheless, another Notice that even if the curve C(t) and its derivatives are
simple representation provides a tighter convex hull as rational in the general case, they become polynomial
well as a better parametrization. In this model, the cir- at the parameter value t = 1because the denominator
cle is composed of six arcs (each of length x / 3 ) and vanishes.
defined by 12 control points (one repeated) on a hexa- Equation 6 shows that at least a continuity exists
gon (see Figure 3c) : at point P,. Moreover, since the two arcs have the same
length, triangles P&Pz and P.&P4 are similar, which
means, in particular, that

Pz-P1= P3-Pz (7)

The junction therefore also has a C1 continuity.The sim-

ilarity of trianglesP&Pz and P&P4, as shown in Figure
5,also provides

i
P4 - P2 = Po - Pz+2 1+cos 24 j (pz Pl 1
-

= Po - Pz + 4w2 0
Pz Pl
- (8)

where even-indexed (respectively, odd-indexed) points The substitution of Equation 8 in the expression of
have their weights set to 1(respectively,&/2). C"(1+) showsthatwecanonlygetC'(l-)=C"(l+) when
Any regular polygon can be used to define the circle, w = 0 (a 180-degree arc, which is forbidden), or w = 1
but the square and the hexagon provide the best trade- (a 0-degree arc, which is not of interest). In other words,
off among cost, accuracy, and ease of use. we cannot define a circle with a C2 continuity using a
quadratic NURBS (or a quadratic piecewise rational
Sphere. Starting from a NURBS-based representa- Bezier) .6,7 Note that w = 0 is allowed when considering
tion of a circle, we can easily obtain a NURBS-basedrep- infinite points, but in that case Equation 6 is not valid
resentation of a sphere by computing a tensor product and there is still a C1 continuityonly (see Piegl and Tiller6
of a set of circles (providingthe parallels) by a set of half- for further details).

66 November 1996
 Note also that the C1 continuity is only obtained in the have to find the actual expression of O ( t ) . For that, we
Euclidian space (that is, after projection). In the associ- involve another classical parametrization of the circu-
ated homogeneous space (before projection), the con- lar arc, the half-tungentparametrization:
tinuity is not better than (? (see Piegl and Tiller6).
vs E [OJ]
Study of the parametrization
Let us take again the circular arc of length 241starting 1-T’s2 2Ts
where T = t a n @ (11)
from the trigonometric origin, illustrated in Figure 2. 1+ T2s2 ’ 1+ T2s2
For many applications, the best parametrization of a
curve is the arc length parametrization (also called uni- The main advantage of this parametrization is that it
form parametrization or constant-speed parametriza- provides a simple relationship between the arc length 8
tion) because the variation of this parameter represents and the parameters:
the exact distance on the curve from one point to anoth-
er. The most obvious need for arc length parametriza-
tion is to provide constant-speed movements in
vs E [OJ]
animation sequences. Such a parametrization is also
useful for surface design or surface rendering; for
instance, arc length parametrization minimizes distor- Therefore, ifwe can find s(t), we will get immediate-
tion when doing texture mapping. ly 0 ( t ) .Fortunately, a fundamental result of projective
In our example, this arc length parametrization is geometry, developed by Moebius,states that two param-
given by the trigonometric parametrization: etrizations of the same curve by a rational polynomial of
the same degree (Equation9 and Equation 11)are relat-
ed by a rational linear function:

a+bt
Combining Equation 3 and Equation 5 gives us the (I
3 ( a , b , c , d ) ~ I R /~ s t =-
c+dt
rational Bezier parametrization of our arc:
Lee and Lucian’’ recently extended this result to piece-
vt E [0,1] wise rational polynomial curves such as NURBS. In our
case, parameters t and s are equal at the boundaries of
(1-2(1- w)t - 2 p - w)t’ the range [OJ] . This induces a more precise expression:
1-2jl-wjt+2(1-wjt2 ’
C(t =j (9)
2Twt -2Tw(l- w)t’

The substitution of Equation 13in Equation 11gives

We can evaluate the quality of this parametrization p = w - 1.Therefore,
by comparing it with the arc length parametrization. In
wt
the ideal case, the two parametrizations are related by
a linear function: (I
st =-
1-t+wt
(14)

vt €[0,1] and, finally,

e(t ) = 24t 7-

t&W2
t r j t ) =2arctan-
Obviously, this is not the case here because it would 1-t+wt
mean that the cosine function could be expressed as a
rational polynomial. When this linear relationship is not The plot of this function, for several values of (I,
fulfilled, we obtain a quantitative evaluation of the appears on the left side of Figure 6 (next page). Note
parametrization by computing the distance between the that the function becomes less straight (in fact, it oscil-
actual function 0(t) and the ideal one 29t. Using a lates around the ideal straight line) when @ increases.
Euclidian norm, this distance A(t) (sometimes called This phenomenon is more apparent on the right side of
chordal deviation) is given by the figure, which shows the deviations of 8(t) from the
ideal horizontal lines. But the best quantitative infor-
mation is given by the chordal deviation A@) plotted in
Figure 7 (next page). This means that except for very
small angles, the representation of a circular arc by a
quadratic NURBS (or a quadratic piece-wise rational
Bezier) involves a parametrization that deviates signif-
Consequently, to evaluate our parametrization, we icantly from the arc length.”

IEEE Computer Graphics and Applications 67

Feature Article

3- enables degree raising while con-

2.5 - trolling precisely the resulting
6 e(t) and W(t) parametrization.
for 4, = 0.25, 0.5,
0.75, and 1 (in Principle
radians). l
0.5 .:m We borrowed our technique’s
general principle from the reparam-
0’
etrization scheme defined by Moe-
0 I I
0 0.5 1 0 0.5 1 bius. Equation 12 showed that the
reparametrization of a rational curve
by a linear rational polynomial mod-
ifies neither its shape nor its degree. A corollary of this
0.04
result is that a reparametrization by a quadratic rational
polynomialdoubles the degree of the initial curve, a cubic
rational polynomialtriples the degree, and so on. There-
0.03
fore, the conic defined by Equation 3, for instance, can be
represented by a quartic rational curve when we replace
7 A(t) for 4, = t by S(t) where Sis a quadratic rational polynomial.
0.02
0.25, 0.5, 0.75, In the most general formulation for S(t), there are six
and 1 (in degrees of freedom:
radians).
0.01 a+bt + c t 2
S(t)=
d+et+ f t 2

0
but several characteristicsappear desirable for this repa-
rametrization. For instance, it is important for the two
parametrizations to have the same domain of variation
( S ( 0 ) = 0 and S(1) = 1).Moreover, due to the isoscelism
1 of the control triangle, the symmetry of the parame-
trization should be conserved (S(1- t) = 1- S(t) or in
other words S’(1- t ) = S’(t)). All these constraints imply
that there remains only one degree of freedom. If we
call this parameter p and define it as the value of the
8 Zigzag derivative at the boundaries (S’(0) = S’(1) = p ) ,
reparametriza- we obtain
tion function 0.5
S(t) for p = 0, pt+(l-p)tZ
0.5, 1,2,6, and S ( t )=
16. 1-2(1- p ) t + 2 p - p)t 2

The shape of this function (see Figure 8) inspired the

name of our reparametrization technique: zigzag repa-
0 rametrization. Notice thatp has to be positive (or even-
tually null) in order to produce an increasing function in
the range [OJ].

zag reparametrization Zigzag reparametrization of a conic

Our discussion has shown that the usual representa- If we apply our zigzag reparametrization on the
tion of circles and spheres by NURBS curves or surfaces Bezier curve defined by Equation 3, we obtain
suffers from two main weaknesses. First, the continuity
Po +2wab PI + b2 P2
of the parametrization is C1 at best. Second, the param- c(t)= a’
etrization can deviate significantlyfrom the arc length. a2 +2wab+b2
In fact, these two weak points have the same origin:
Because the degree of the curve is only quadratic, there with a=l-(2-p)t+(l-p)t2 and b=pt+(1-p)t2. The
are not enough degrees of freedom to correct one (or result is a quartic rational curve, but it is not a rational
both) faults. We might therefore raise the degree to get Bezier since it is no longer written in terms of Bernstein
more parameters to manipulate. Unfortunately, the clas- polynomials. But because the Bernstein polynomials
sicalprocess of raising the degree of a Bezier or a B-spline form a basis for the polynomial space, we are assured
without modifying its shapeZ8,l3does not work for this that there exists a Bezier curve equivalent to Equation
particular goal because it does not change the parame- 17. To find this quartic rational Bezier curve, we have to
trization. For that reason, we propose here a technique find five points (Qo,QI, Q2, Q3, Q4) and five weights (WO,
(restricted to rational or piece-wiserational curves) that W I ,W Z ,w3, wq) that obey Equation 18 (Figure 9 ) .

68 November 1996
 ~ , ( l - t ) ~+Q4 ~ , t ( l - t ) ~ Q+6W2t2(1-trQ,
, +4W3t3(l-f) Q 3 +w4t4Q4 9 Equation 18.
C(t)=
wo(l- t )4 +4w,t(l- t )3 + 6w2t2(1-t)’ +4w,t3(1- t ) + w4t4

Obviously, QO= Po and Q4 = PZbecause of the inter- extension for the reparametrization of nonsymmet-
polation of the boundaries. We also know that we can ric curves such as ellipses or parabolas.
let wo = 1 and w4 = 1for symmetry reasons, without Zigzag reparametrization can also be extended above
loss of generality.We obtain the remaining weights wlr degree 2. We do this by defining a family of rational
w2, and w3 by comparing the denominators of Equation polynomials where each member Sk(t) of degree k
18 (see Figure 9) and Equation 17 term by term. fulfills
Similarly, we obtain the remaining points QI, Qz,
and Q3
by comparing the numerators.
Finally, the zigzag reparametrization of Equation 17
provides a quartic rational Bezier (and therefore quar-
tic NURBS) defined by
3p, E I R + / -d‘S
(O)=~,
Qo =Po
dt‘

0
$Po +2w l + p 2 P1+p2P2
Therefore,each successivemember of this zigzagpoly-
Qz = 2jpz +p2w+w
i nomialfamily includes one additional parameter that
enables more precise control of the parametrization.

Next, we show how to use zigzag reparametrization

Q4 =P2
to improve the parametrization of a circle represented by
wg =1 a NURBS. We provide these examples only as illustra-
l+w tion. The new technique may in fact be employed in
w1 = -p many applications that involve reparametrization
2
according to a given criterion.
+p2w+w
p2
w2 =
3 A C2 parametrization of the circle
l+w We first apply zigzag parametrization to obtain a cir-
w3 =- P
2 cle with a C2parametrization. Starting again from the
w4 =1 configuration of control points illustrated in Figure 5,
we can study the continuity at point P2:

Four important observations emerge from this for-

mulation:

W Only the position of point Q 2 depends on the value of

parameterp. A nice consequence of this is that the
zigzag reparametrization has a very simple geomet-
ric interpretation: changing the value ofp consists in
moving Qz along a straight line (more precisely, a half
line because we forcep 2 0). The simultaneous mod-
ification ofwl, w2, and w3 keeps the curve unchanged
(except for its parametrization) despite the displace-
ment of Qz.
H Whenp = 1,we obtain exactlythe points and weights As previously, Equation 7 yields C1 continuity. To
provided by the classical degree raising technique check the C2continuity, we have to substitute Equation
mentioned above. In other words, it means that 7 and Equation 8 in the expression of C“(1’):
zigzag reparametrization is a kind of generalizationof
this process, which provides an additional degree of
freedom.
The principle of the zigzag reparametrization can be
extended to nonsymmetric reparametrization by per- Finally, comparing C”(1-) and C”(1’) provides
mitting two degrees of freedom ( p = S’(0) and
q = S’(1)) for Equation 16. We can also use such an p2(1+ w)- p -1 = 0 (21)

IEEE Computer Graphics and Applications 69

I C L l L u l c _I C 3 L . L

Q3 same length, with the resulting curve being represent-

ed either by quartic rational Bezier or a quartic NURBS.
To illustrate this, Figure 10 shows a circle with a C2
10 C’ repara- Qi parametrization obtained by zigzag reparametrization
metrization of of the triangle-based circle, using for p the value pro-
the triangle- QO
vided by Equation 22.
based circle.

A second application of the zigzag reparametrization

is to obtain a circle with a quasi-uniform parametriza-
tion, as close as possible to the trigonometric parame-
trization. As we have seen, the parametrization error
can be quantified by computing the chordal deviation
0.0004 A(t) between the arc length parametrization and the

;=,
actual one. Substituting Equation 16 in Equation 10
expresses the chordal deviation for a circular arc on
0.0003
which we have applied the zigzag reparametrization.
11 A(t)for$= The formulation of this new chordal deviation is a func-
0.25, 0.5,0.75, 0.0002 tion of parameterp:
and 1 (in
radians).
2 arctan
& Z q p t + (1-p)tZ) - 24t
0.0001
A(t) =
I-%+
i
I+W
i pt+ I - p t
2, (23)
n
0 0.2 0.4 0.6 0.8 1
JZ
What we have here is a classical
optimizationproblem: Find the opti-
3 mal value of a parameter (in our
12 e(t) and 2.5 case, the reparametrization factorp)
W(t) for 4 = 2 that minimizes a given condition (in
0.25,0.5, 0.75, our case, the chordal deviation A(t)).
1.5
1, 1.25, and 1.5 One possibility is to minimize the Lz
1 1 norm of A(t) by employing a least-
(in radians).
0.5 0.5 squares minimization method. The
0 0 drawback of this technique (and all
0 0.5 1 0 0.5 other techniques of the same fami-
ly) is that it provides only a numeri-
cal (and not an analytical) solution. Thus, for each new
Q3 angle I$needed, the user has to restart the whole least-
13 Quasi- squares process to find the new parameterp.
uniform Another possibility is to minimize theLwnorm of A(t),
reparametriza- Qi which only requires computing the maximal chordal
tion of the deviation. This maximum has an analytical expression
triangle-based Qo that can be computed by symboliccalculation software
circle. such as Maple or Mathematica, but this expression is so
complex as to be useless in practice.
For these reasons, we propose here a third solution
that is only heuristic but, in fact, very close to the opti-
mum. If we study the variation of the maximal chordal
deviation for different values of @, we can see that this
maximum always results for a value t that belongs in
This quadratic equation has always two solutions [0.19, 0.241. If we take a value in that range ( t = 1/5
(A=4w+5>0), but only the positive solution is useful seems to work well), we are assured that the chordal
for our reparametrization: deviation at point A ( l / 5 ) will not be too far from the
maximal deviation. Therefore, we simply have to force
A(l/5) = 0, that is,
1+45+ 4w
P=
2(1+ w) (I+ 4p)sin4
- tan(4/5) = o (24)
Therefore, Equation 22 combined with Equation 19 (1
18+ 12p - + 4p)cos$
enables a C2continuous junction for circular arcs of the

70 November 1996
to get a quasi-optimal value forp:

4-2cos3 (@/5)+cos(@/5)
P= (25)
-1+ 8c0s3(@/5)--4cos($/5)

Figure 11shows the plot of the chordal deviation ( t )

obtained with the quasi-optimal value p . Notice that
there is a factor of 100in the scale between Figure 7 and
Figure 11,meaning that the parametrization error has
been reduced more than 100 times. Figure 12 confirms
this result, showing the variations of O(t) and O’(t) for
different values of $ when using the quasi-optimal para-
meter: The resulting curves are almost straight, and no
more oscillations are visible.
Figure 13 illustrates the circle with a quasi-uniform
parametrization obtained by zigzag reparametrization
of the triangle-based circle, using the parameter p pro-
vided by Equation 25. Comparing Figure 10 and Figure
13 shows that the C2 continuous circle and the quasi-
uniform circle are almost identical. In fact, the differ-
ence between the reparametrizations is so small
(p 1.21525 for the first andpz 1.20669 for the second) ing Surfaces of Revolution Using Boundary Data,” IEEE
that the variation of the position of the control points is CG&A, Vol. 7, No. 3, May 1989, pp. 45-55.
hardlyvisible (remember that only Qz, Q6, and Qlo have 9. W. Tiller, “RationalB-Splines for Curve and Surface Repre-
actually moved). sentation,”IEEECG&A,Vol. 3, No. 6, Sept. 1983,pp. 61-69.
Table 1presents all the numerical results relative to 10. C. Blanc, Mod&ation etD&ormation des Surfacespour la
zigzag reparametrization of the triangle-based, square- Synthise d’lmages, doctoral dissertation (in French), Uni-
based, and hexagon-based circles. Notice that the versite de Bordeaux, Bordeaux,France, 1994.
chordal deviation can be reduced more than 1,000times 11. E. Lee and M. Lucian, “Moebius Reparametrization of
by the reparametrization. Notice also that the analytic RationalB-Splines,”ComputerAided GeometricDeszgn,Vol.
LM minimizationoffers only an average of 5 percent addi- 8,1991,pp. 213-238.
tional precision compared to our heuristic minimization, 12. R. Farouki and T. Sakkalis, “Real Rational Curves are not
for a much more expensive computation cost. This con- Unit Speed,”ComputerAided GeometricDesign, Vol. 8, No.
firms the nice behavior of the “A(1/5) = 0 heuristic. 2,1991,pp. 151.158.
To conclude, we would like to note that we have cho- 13. R. Bartels, J. Beatty, and B. Barsky, An Introduction to
sen to use the circle here only as an illustrating exam- Splines for Computer Graphics and Geometric Modeling,
ple. The idea of zigzag reparametrization is much more Morgan Kaufmann, San Francisco, 1987.
general and we feel that it could benefit many situations
where well-behaving parametrization is important. W
C a r o l e B l a n c is a lecturer at the
Universityof Bordeaux, France, where
she received a PhD in Computer Sci-
References ence in 1994. Her researchfocuses on
1. L. Piegl and W. Tiller, TneBook ofNURsS, SpringerVerlag, geometric modeling, includingpara-
New York, 1995. metric surfaces, implicitsurfaces, and
2. G. Farin, NURB Curves and Surfaces, A.K. Peters, Boston, deformation techniques.
1995.
3. A. Forrest, Curves and Surfaces f o r Computer-AidedDesign,
doctoral dissertation, Cambridge University, Cambridge, ChristopheSchlickisan assistant
U.K., 1968. professor at the University of Bor-
4. K. Versprille, Computer Aided Design Applications of the deaux, where he obtained a PhD in
Rational &Spline Approximation Form, doctoral disserta- Computer Science in 1992. His
tion, University of Syracuse,New York, 1975. research interests include illumination
5. G. Farin, “FromConics to NURBS: ATutorial and Survey,” models, Monte Carlo techniques,pro-
IEEE Computer Graphics &Applications, Vol. 12, No. 5, cedural textures, and spline curves
1989,pp. 78-86. and surfaces.
6. L. Piegl and W. Tiller, “A Menagerie of Rational B-Spline
Circles,”IEEE C G M , Vol. 9, No. 5, Sept. 1989,pp. 48-56. Contact the authors at the Laboratoire Bordelais de
7. G. Farin, Curves and Surfaces f o r Computer Aided Geomet- Recherce en Informatique, University of Bordeaux, 351
ricDesign, Academic Press, 1990. Cours de la Liberation, 33405 Talence, France, e-mail
8. L. Piegl, “InfiniteControl Points: AMethod for Represent- {blanc, schlick)@labri.u-bordeaux.fi.

IEEE Computer Graphics and Applications 71