NURB Curves: A Guide for the Uninitiated

Philip J. Schneider

QuickDraw 3D supports a mathematical model for arbitrary curves and surfaces known as NURB (nonuniform rational B-splines). NURB curves are flexible and powerful, but using them effectively requires some understanding of the underlying mathematical theory. This article presents an intuitive introduction to the mathematical concepts of the NURB model and how to use them in your QuickDraw 3D programs.

One of the more powerful features of QuickDraw 3D is its ability to work with curves and surfaces of arbitrary shape. The mathematical model it uses to represent them is known as NURB, for nonuniform rational B-splines. The NURB model is flexible and powerful, but for those unfamiliar with the mathematics, it can appear dauntingly complex. The existing books and articles on the subject tend to be rigorous, lengthy, and theoretical, and often seem to require that you already understand the subject in order to follow the explanations.

The mathematics really aren't so frightening, though, once you understand them. The aim of this article is to give you an intuitive understanding of how NURB curves work. Later in the article, we'll look at some code to show you how you can start using NURB curves in your own programs -- but you really do need to understand the theory before you can start putting it to practical use. So please be patient while we slog through the mathematical concepts: I promise we'll get around to some actual programming before we're through. Note also that this article is only about NURB curves; perhaps a future article will cover NURB surfaces and how to use curves and surfaces together.

Some writers also use the s from "spline," resulting in the acronym NURBS -- but most avoid this usage because phrases like "a NURBS curve" sound awkward, and "a NURBS surface" sounds perfectly hideous.*

WHY NURB CURVES?

Like any graphics package, QuickDraw 3D offers low-level geometric primitives for objects such as lines, points, and triangles. Because the representations of these objects are mathematically exact -- lines being defined by their two endpoints, triangles by their three vertices, and so forth -- they're resolution independent and unaffected by changes in position, scale, or orientation.

The low-level primitives can also be used to define arbitrarily shaped objects, such as a football or an automobile hood, but at the cost of these desirable mathematical properties; for example, a circle that's approximated by a sequence of line segments will change its shape when rotated. One of the advantages of NURB curves is that they offer a way to represent arbitrary shapes while maintaining mathematical exactness and resolution independence. Among their useful properties are the following:

• They can represent virtually any desired shape, from points, straight lines, and polylines to conic sections (circles, ellipses, parabolas, and hyperbolas) to free-form curves with arbitrary shapes.
• They give you great control over the shape of a curve. A set of control points and knots, which guide the curve's shape, can be directly manipulated to control its smoothness and curvature.
• They can represent very complex shapes with remarkably little data. For instance, approximating a circle three feet across with a sequence of line segments would require tens of thousands of segments to make it look like a circle instead of a polygon. Defining the same circle with a NURB representation takes only seven control points!

In addition to drawing NURB curves directly as graphical items, you can use them in various other ways that exploit their useful mathematical properties, such as for guiding animation paths or for interpolating or approximating data. You can also use them as a tool to design and control the shapes of three-dimensional surfaces, for purposes such as

• surfaces of revolution (rotating a two-dimensional curve around an axis in three-dimensional space)
• extruding (translating a curve along a curved path)
• trimming (cutting away part of a NURB surface, using NURB curves to specify the cut)

CURVES 101

Before we go into the specifics of NURB curves, let's review some of the basics of curve representation in general.

Although QuickDraw 3D supports three-dimensional NURB curves, we'll limit all of our examples and discussions here to two dimensions. But everything we say about two-dimensional curves applies in three dimensions as well -- the two-dimensional versions are just easier to visualize and easier to draw.

A BIT OF HISTORY

Back in the days before computers, architects, engineers, and artists would draw their designs for buildings, roads, machine parts, and the like by using pencil, paper, and various drafting tools. These tools included rulers and T-squares for drawing straight lines, compasses for drawing circles and circular arcs, and triangles and protractors for making precise angles.

Of course, a lot of interesting-shaped objects couldn't be drawn with just these simple tools, because they had curved parts that weren't just circles or ellipses. Often, a curve was needed that went smoothly through a number of predetermined points. This problem was particularly acute in shipbuilding: although a skilled artist or draftsman could reliably hand-draw such curves on a drafting table, shipbuilders often needed to make life-size (or nearly life-size) drawings, where the sheer size of the required curves made hand drawing impossible. Because of their great size, such drawings were often done in the loft area of a large building, by a specialist known as a loftsman. To aid in the task, the loftsman would employ long, thin, flexible strips of wood, plastic, or metal, called splines. The splines were held in place with lead weights, called ducks because of their resemblance to the feathered creature of the same name (see Figure 1).

Figure 1. A draftsman's spline

The resulting curves were smooth, and varied in curvature depending on the position of the ducks. As computers were introduced into the design process, the physical properties of such splines were investigated so that they could be modeled mathematically on the computer.

DIRECT FUNCTIONS

Our goal is to represent curves in a mathematically precise fashion. One simple way is to think of the curve as the graph of a function:

y = f(x)

Take a simple one like the trigonometric sine function:

y = sin x

By plotting the value of the function for various values of x and connecting them smoothly, we obtain the curve shown in Figure 2.

Figure 2. Plot of sine function values

In the case of curves drawn by the spline method, it turned out that with some reasonable simplifying assumptions, they could be mathematically represented by a series of cubic (third-degree) polynomials, each having the form

y = Ax³ + Bx² + Cx + D

At this point, the standard references typically go into a long, involved development of this idea into what are known as cubic spline curves, eventually leading to the theory of NURB curves. Such explanations are interesting, but not terribly intuitive. If you're interested in pursuing this subject further, you'll find a good discussion in Mathematical Elements for Computer Graphics. (Complete information on this and other literature references in this article can be found in the bibliography at the end.)

PARAMETRIC FUNCTIONS

Using direct functions to represent a curve fits our criterion of being mathematically exact, but it has one serious drawback: since we can have only one value of y for each value of x, our curves can't loop back on themselves. Thus, although we can make some nice smooth curves this way, there are a lot of interesting curves we can't make -- not even something as simple as a circle.

An alternative method, and the one we'll be using, is to define the curve with a parametric function. In general, such functions have the form

Q(t) = {X(t), Y(t)}

where X(t) and Y(t) are functions of the parameter t (hence parametric). Given a value of t, the function X(t) gives the corresponding value of x, and Y(t) the value of y. One way to understand such functions is to imagine a particle traveling across a sheet of paper, tracing out a curve. If you think of the parameter t as representing time, the parametric function Q(t) gives the {x, y} coordinates of the particle at time t. For example, defining the functions X(t) and Y(t) as

X(t) = cos t

Y(t) = sin t

produces a circle, as you can verify by plugging in some values of t between 0 and 2[[pi]] and plotting the results.

SMOOTHNESS

One very important motivation for using NURB curves is the ability to control smoothness. The NURB model allows you to define curves with no kinks or sudden changes of direction (such as an airplane-wing cross section) or with precise control over where kinks and bends occur (sharp corners of machined objects, for instance).

We all know (or think we do) what a nice, smooth curve looks like: it has no kinks or corners. If we were to sit on that moving particle as it traces out a parametric curve, we would experience a nice smooth ride with no stopping, restarting, or sudden changes in speed or direction: we wouldn't be heading north, say, and then turn completely east in an instant. This intuitive notion can be expressed in precise mathematical terms: Imagine an arrow that always points in the direction in which our hypothetical particle is traveling as it moves along the curve. Mathematically, the direction arrow corresponds to the tangent of the curve, which can be computed as the derivative of the curve's defining function with respect to the time parameter t:

Q'(t)

In Figure 3, for example, the point on the curve corresponding to time tA is labeled as Q(tA), and the direction vector (tangent) at that point as Q'(tA). If the tangent doesn't jump suddenly from one direction to another, the curve's function is said to have first-derivative continuity, denoted by C¹: this corresponds to our intuitive notion of smoothness.

Figure 3. Tangent (derivative) of a curve

Now look at the point marked Q(tB), where there's a visible kink in the curve. The direction vector just a tiny bit to the left of that point, Q'(tB-[[alpha]]), is wildly different from the one just a tiny bit to the right, Q'(tB+[[alpha]]). In fact, the direction vector jumps instantaneously from one direction to another at point Q(tB). Mathematically, this is called a discontinuity.

Many of you will recall from your college calculus that the derivative of a function is also a function, whose degree is one less than that of the original function. For example, the derivative of a fourth-degree function is a third-degree function. The derivative of the derivative, called the second derivative, will then be of degree 2. This second derivative may or may not be continuous: if it is, we say that the original function has second-derivative continuity, or C². As the first derivative describes the direction of the curve, the second derivative describes how fast that direction is changing. The second derivative thus characterizes the curve's degree of curvature, and so a C²-continuous curve is said to have curvature continuity. We'll come back to these important concepts after we've introduced NURB curves themselves.

NURB CURVES

Now that we know how parametric functions work, let's see how we can use them to build up a definition for NURB curves. If we call our function Q, the left side of our equation will look like this:

Q(t) =

where t is a parameter representing time. By evaluating this function at a number of values of t, we'll get a series of {x, y} pairs that we can use to plot our curve, as shown in Figure 4. Now all we have to do is define the right-hand side.

Figure 4. Plotting a parametric function

CONTROL POINTS

One of the key characteristics of NURB curves is that their shape is determined by (among other things) the positions of a set of points called control points, like the ones labeled Bi in Figure 5. As in the figure, the control points are often joined with connecting lines to make them easier to see and to clarify their relationship to the curve. These connecting lines form what's known as a control polygon. (It would actually make more sense to call it a "control polyline," but the other is the conventional term.)

Figure 5. Defining a curve with control points

The second curve in Figure 5 is the same curve, but with one of the control points (B7) moved a bit. Notice that the curve's shape isn't changed throughout its entire length, but only in a small neighborhood near the changed control point. This is a very desirable property, since it allows us to make localized changes by moving individual control points, without affecting the overall shape of the curve. Each control point influences the part of the curve nearest to it but has little or no effect on parts of the curve that are farther away.

One way to think about this is to consider how much influence each of the control points has over the path of our moving particle at each instant of time. At any time t, the particle's position will be a weighted average of all the control points, but with the points closer to the particle carrying more weight than those farther away. We can express this intuitive notion mathematically this way:

In other words, to find the position of the moving particle at a particular time, add up the positions of all the control points (Bi) but vary the strength of each point's contribution over time (Ni,k(t)). We'll explain the meaning of the subscript k shortly.

The bounding volume returned by QuickDraw 3D for all other geometric primitives is a volume that encloses the primitive itself. For NURB curves, however, the returned bounding volume encloses the curve's control points, rather than the curve itself. This is done for historical reasons, and is the normal practice in 3D graphics packages.*

BASIS FUNCTIONS

The function Ni,k(t), which determines how strongly control point Bi influences the curve at time t, is called the basis function for that control point. In fact, the B in "B-splines" stands for "basis." The value of this function is a real number such as 0.5, so that a particular point Q(t) can be defined as, say, 25% of one control point's position, plus 50% of another's, plus 25% of yet a third's. To complete our NURB equation, we have to specify the basis function for each control point.

So how do we go about defining the basis functions? Remember that we want each region of the NURB curve to be a local average of some small number of control points close to that region. When the moving particle is far away from a given control point, that control point has little influence on it; as the particle gets closer, the control point affects it more and more. Then the effect diminishes again as the particle recedes past the control point.

Up to now, we've been using the words "near" and "far" in a rather vague way, but the time has come to pin them down more rigorously. Because we've defined our curve parametrically with respect to time, we can regard what we've been calling a "part" or "region" of the curve as a portion of the time interval the curve covers. For example, if our curve goes from time t = 0.0 to t = 10.0, we can specify a particular region as extending from, say, t = 3.3 to t = 7.5. So we can say, for instance, that a control point Bi is centered at time t = 5.0 and has an effect in the range from t = 3.3 to t = 7.5.

Figure 6 shows a typical example of what a basis function might look like: it has its maximum effect at some definite point in time and tapers off smoothly as it gets farther away from that point. If you were awake during your college statistics course, you might recognize this as the familiar "bell curve" that we all learned to know and loathe. The curve Ni,k(t) in the figure shows that control point Bi has its greatest effect (about 95%) at time t = 3.0 and tapers off to about 50% at t = 1.7 and t = 4.3.

Figure 6. Basis function for a control point

Since each control point has its own basis function, a NURB curve with, say, five control points will have five such functions, each covering some region of the curve (that is, some interval of time). At time t = 2.3 in Figure 7, for example, control point B0 has a weight of about 0.2, B1 about 0.7, and B2 about 0.05. As t goes from 0.0 to 7.0, each control point's effect on the shape of the curve is initially 0, increases gradually to a maximum, and then gradually tapers off again to 0 as we reach the end of its effective range.

Figure 7. Uniform basis functions for a set of control points

KNOTS

Notice that all of the basis functions in Figure 7 have exactly the same shape and cover equal intervals of time. In general, we'd like to be able to vary the width of the intervals (so that some control points affect a larger region of the curve and others a smaller region) and the maximum height of the curves (so that some control points affect the shape of the curve more strongly than others). That's where the NU in NURB comes from: it stands for nonuniform.

The solution is to define a series of points that partition the time into intervals, which we can then use in the basis functions to achieve the desired effects. By varying the relative lengths of the intervals, we can vary the amount of time each control point affects the particle. The points demarcating the intervals are known as knots, and the ordered list of them is a knot vector (Figure 8). The knot vector for the basis functions shown in Figure 7 is {0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0}. This is an example of a uniform knot vector, which is why all the functions in the figure cover equal intervals of time. Figure 9 shows an example of a curve created with such a knot vector.

Figure 8. A knot vector

Figure 9. NURB curve with uniform knot vector

If we change the knot vector to {0.0, 1.0, 2.0, 3.75, 4.0, 4.25, 6.0, 7.0}, we get a set of nonuniform basis functions like the ones shown in Figure 10, and a curve that looks like Figure 11 (using the same set of control points as in Figure 9). Notice that the basis functions N2,3(t) and N3,3(t), associated with control points B2 and B3, respectively, are taller and narrower than the others. If you compare Figures 9 and 11, you'll see that the curve in Figure 11 is pulled more strongly toward control points B2 and B3 than the one in Figure 9. This is because the basis functions for these control points have a greater maximum value. Also, the curve rapidly approaches these control points and rapidly moves away: compare how tightly curved it is near these points, relative to the curve in Figure 9. This is a result of the narrower basis functions for these two control points: intuitively, our moving particle has to traverse more space in relatively less time. Looking at the knot vector, you can see that the knot intervals for these two control points are narrower than the others -- {3.75, 4.0} and {4.0, 4.25} -- meaning that their effects on the curve are concentrated in shorter time intervals.

Figure 10. Nonuniform basis functions for a set of control points

Figure 11. NURB curve with nonuniform knot vector

DEFINING THE BASIS FUNCTIONS

We're now ready to complete our definition of a NURB curve by giving an exact specification of the basis functions. In some respects, we're free to use any sort of functions we'd like, but by choosing them carefully, we can get certain desirable effects. The definitions we'll be using are as follows:

where xi is the conventional notation for the ith knot in the knot vector.

This definition has a lot of stuff in it, and lots of subscripts -- we're getting into the real theoretical aspects of NURB curves here. Notice that the functions for higher values of the subscript k (called the order of the basis function) are built up recursively from those of lower orders. If k is the highest order of basis function we define, the resulting NURB curve is said to be of order k or of degree k-1. At the very bottom of the hierarchy, the functions of order 1 are simply 1 if t is between the ith and (i+1)st knots, and 0 otherwise.

The specifics of this particular set of basis functions, and how they came to be this way, are beyond the scope of this article; if you're interested in learning more, you'll find all the detail you could possibly want (and then some) in An Introduction to Splines for Use in Computer Graphics and Geometric Modeling. However, we can at least mention a number of important characteristics that this choice of basis functions exhibits:

• At any time t, the values of all the basis functions add up to exactly 1.
• If all control points have positive weights, the curve is contained within a bounding region known as the convex hull. (See the book cited above for details.)
• At any time t, no more than k basis functions affect the curve, where k is the order of the curve.
• A curve of order k is defined only where k of the basis functions are nonzero.

This last characteristic is of more than theoretical interest: a cubic (degree-3 or order-4) NURB curve with a knot vector of, say, {0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0} only goes from t = 3.0 to t = 4.0! The rule is that the curve begins at the kth knot from the beginning of the knot vector and ends at the kth knot from the end.

KNOTS AND KINKS

I should point out here that nonuniform knot vectors aren't really very useful for controlling the shape of a curve. (In fact, moving control points around directly isn't that useful, either -- but we'll get to that later.) Instead, nonuniform knot vectors have two important uses:

• You've probably noticed that all of our NURB curves so far have had their endpoints just "floating in space"; that is, the curve's endpoints don't coincide with any control point. In real life, though, we generally want to be able to control the exact placement of the endpoints, and most often we want them to coincide exactly with the first and last control points.
• You may also have noticed that the curves displayed so far are quite smooth. While this is usually a good thing, we sometimes need to create a curve with a kink or corner.

We can accomplish both of these goals by using a rather extreme case of nonuniformity: giving several consecutive knots the same value of t! For example, a knot vector like {0.0, 0.0, 0.0, 3.0, 4.0, 5.0, 6.0, 7.0} produces a set of basis functions like those in Figure 12 and a curve (using the same control points as before) that looks like Figure 13. Looking at Figure 12, you can see that at t = 0, the basis functions associated with all but the first control point have a 0 value -- so basis function N0,3(t) (the one for control point B0) has total control over the curve. Thus the curve at t = 0 coincides with the first control point.

Figure 12. Basis functions for a curve with multiple identical knots at the beginning

Figure 13. NURB curve with multiple identical knots at the beginning

If we bunch up some knots in the middle of the knot vector {0.0, 1.0, 2.0, 3.0, 3.0, 5.0, 6.0, 7.0}, we get the basis functions shown in Figure 14 and the curve in Figure 15. At t = 3.0, all the basis functions except N2,3(t) have a 0 value -- so control point B2 is the only one to affect the curve at that instant, and thus the curve coincides with that control point.

Figure 14. Basis functions for a curve with multiple identical knots in the middle

Figure 15. NURB curve with multiple identical knots in the middle

In mathematical terms, continuity (smoothness) is an issue only at the joints defined by the curve's knots, where two segments of the curve meet; between the joints, the curve is perfectly smooth and continuous. A typical curve, in which each joint corresponds to a single knot, has continuity C^n-1 where n is the degree of the curve. So a cubic (degree-3 or order-4) curve has second-derivative continuity (C²) at each joint if all the knots are distinct. If two knots coincide, the continuity at that joint goes down by one degree; if three coincide, the continuity goes down another degree; and so on.

This means you can put a kink in the curve at a particular point by adding knots to the knot vector at that point. Later, we'll look at some code that shows how to do this. We'll also see how you can use this same technique of knot insertion to convert a curve from NURB to Bézier representation.

RATIONAL CURVES

Now that we've learned all about control points and knots and basis functions, we understand NUB (nonuniform B-spline) curves. But what about the rest of the acronym? We're still missing the R in NURB. It's time to talk about rational curves.

If you've sneaked a peek at QuickDraw 3D's NURB definitions, you may have wondered why it uses a four-dimensional representation for three-dimensional control points: {x, y, z, w} instead of just {x, y, z}. The reason for the extra coordinate is that it allows us to exactly represent conic curves (circles, ellipses, parabolas, and hyperbolas), as well as giving us more control over the shape of other curves. The fourth coordinate, w, is customarily referred to as the weight of the control point. Ordinarily, each control point carries a weight of 1.0, meaning that they all have equal influence on the shape of the curve. Increasing the weight of an individual control point gives it more influence and has the effect of "pulling" the curve toward that point (see Figure 16).

Figure 16. Increasing the weight of a control point

Curves that are defined in this way, with a weight w for each control point, are called rational curves. Mathematically, such curves are defined in four-dimensional space (since the control points have four components) and are projected down into three-dimensional space. Visualizing objects in four dimensions is a bit difficult (let alone drawing them in a diagram), but we can understand the basic idea by considering rational two-dimensional curves: that is, curves defined in three-dimensional space and projected onto a plane, as shown in Figure 17.

Figure 17. Projecting a three-dimensional curve into two dimensions

This is essentially the same process as projecting a three-dimensional model onto a two-dimensional screen with a perspective camera. The basic method for such perspective projection is to divide by the homogeneous component of the vertex (that is, w); we use an analogous approach to project our four-dimensional rational curve into three-dimensional space. Mathematically, then, we must incorporate this division into our earlier definition for a B-spline curve:

The Bi are the projections of the four-dimensional control points and the wi are their weights.

There are two different conventions for representing the control points in terms of their four-dimensional coordinates {x, y, z, w}:

• Homogeneous, in which the coordinates represent the point's position in four-dimensional space. To project it into three dimensions, the components must all be divided through by w. Thus the point's three-dimensional position is actually {x/w, y/w, z/w}. (Note that w/w is always 1.)
• Weighted Euclidean, in which the coordinates are already considered to have been divided through. Thus the first three components {x, y, z} directly represent the point's position in three-dimensional space and the fourth (w) represents its weight.

QuickDraw 3D uses homogeneous representation, as do most technical papers and other graphics libraries.

CONIC SECTIONS

I said earlier that we could use the rational aspect of NURB curves to create conic sections (such as circles and ellipses). Conic sections are so called because they're the curves we get by intersecting a cone with a plane; the angle at which the plane intersects the cone determines whether the resulting curve is a circle, an ellipse, a parabola, or a hyperbola. Strictly speaking, hyperbolas and parabolas are of infinite extent -- but infinite curves are generally not useful in graphics applications (besides being very hard to compute a bounding box for). So we'll restrict our discussion to conic arcs.

Since conic curves are quadratic, we can represent them by quadratic (degree-2 or order-3) NURB curves. The practical question, of course, is which NURB curve. Although the proof is beyond the scope of this article, the following method (illustrated in Figure 18) can be used to generate conic arcs:

Figure 18. Constructing conic arcs

• The curve is defined by three control points. The first and last are the endpoints of the conic arc, while the placement of the inner control point helps determine the shape of the curve.
• The weights of the first and last control points are 1.0.
• A weight less than 1.0 for the inner control point generates an ellipse; a weight equal to 1.0 generates a parabola; a weight greater than 1.0 generates a hyperbola.
• The knot vector is {0.0, 0.0, 0.0, 1.0, 1.0, 1.0}.

Probably the most common form of conic arc, particularly in modeling and design applications, is a circular arc. Since a circle is simply a special case of an ellipse, the method for constructing a circular arc is a special case of the general method for elliptical arcs:

• The legs of the control polygon are of equal length (that is, the control triangle is isosceles).
• The chord connecting the first and last control points meets each leg at an angle [[theta]] equal to half the angular extent of the desired arc (for instance, 30deg. for a 60deg. arc).
• The weight of the inner control point is equal to the cosine of [[theta]].
• The knot vector is {0.0, 0.0, 0.0, 1.0, 1.0, 1.0}, just as before.

Figure 19 illustrates this construction. (In this case, the control triangle is equilateral, so the angle [[theta]] is 60deg. and the resulting arc is 120deg., or one-third of a circle.)

Figure 19. Constructing a circular arc

Note that the foregoing method can only produce circular arcs less than 180deg.; for larger arcs, we have to piece together several NURB curves. So to draw a complete circle we could combine three 120deg. arcs, or four 90deg. arcs. However, it's possible to represent these three or four separate arcs as a single curve and to make a circle with only one NURB curve. Figures 20 and 21 show how to do it with three and four arcs, respectively.

Figure 20. Constructing a circle with three arcs

Figure 21. Constructing a circle with four arcs

NURB CURVES IN QUICKDRAW 3D

By now you're probably saying, "Enough theory already -- how does all this relate to Macintosh programming?" So let's finally look at QuickDraw 3D's data structures and routines for working with NURB curves.

DATA STRUCTURES

If you've been following the discussion so far, you can probably guess the contents of the data structure representing a NURB curve: the order of the curve, its control points, and its knots. There's also the usual QuickDraw 3D attribute set, so you can draw your curves in, say, fuchsia or vermilion. Here are the definitions:

typedef struct TQ3RationalPoint4D {
   float      x;
   float      y;
   float      z;
   float      w;
} TQ3RationalPoint4D;

typedef struct TQ3NURBCurveData {
    unsigned long       order;            // Order of the curve
    unsigned long       numPoints;        // Number of control points
    TQ3RationalPoint4D  *controlPoints;   // Array of control points
    float               *knots;           // Array of knots
    TQ3AttributeSet     curveAttributeSet; // QuickDraw 3D attributes
} TQ3NURBCurveData;

Most of this is pretty straightforward, but here are a few things to keep in mind:

• The order of the curve must be between 2 and 16, inclusive. Order 2 gives you a polyline effect; the most common orders are 3 (quadratic) and 4 (cubic).
• The control points are represented in homogeneous form, meaning that you have to divide the x, y, and z components by the w component to find the point's actual position in three-dimensional space.
• The w component of each control point must be positive.
• The number of control points must be equal to or greater than the order.
• The number of knots must be equal to the number of control points plus the order of the curve.
• The knots must be specified in nondecreasing order.
• If k is the order of the curve, there can't be more than k-1 knots with the same value (except at the beginning or end of the sequence, where k consecutive equal knots are allowed).
• The attribute set should contain only attributes that make sense for a curve. Most often, the attribute set will either be NULL or simply contain a color.

RENDERING A NURB CURVE

If you're familiar with QuickDraw 3D, you know that there are two ways to render a graphical entity (called a geometry in QuickDraw 3D terminology): retained mode and immediate mode. In retained mode, you first create an object representing the figure you want to draw, then use this retained object to do your drawing. (See the article "The Basics of QuickDraw 3D Geometries" in develop Issue 23 for more on this.) Listing 1 shows how this works for a NURB curve. First we initialize a TQ3NURBCurveData structure describing the curve to be drawn; we use this structure to create a retained object with the QuickDraw 3D function Q3NURBCurve_New, and then we pass the resulting object to Q3Geometry_Submit to render the curve. Finally, we dispose of the retained object we created.

Listing 1. Rendering a NURB curve in retained mode

TQ3GeometryObject         curveObject;
TQ3NURBCurveData         curveData;

static TQ3RationalPoint4D  controlPoints[4] = {
                             { 0, 0, 0, 1 },
                             { 1, 1, 0, 1 },
                             { 2, 0, 0, 1 },
                             { 3, 1, 0, 1 }
                           };
static float               knots[8] = {
                              0, 0, 0, 0, 1, 1, 1, 1
                           };

// Initialize the data structure.
curveData.order               = 4;
curveData.numPoints           = 4;
curveData.controlPoints       = controlPoints;
curveData.knots               = knots;
curveData.curveAttributeSet   = NULL;

// Make a retained object.
curveObject = Q3NURBCurve_New(&curveData);

// Use the retained object to render the curve.
Q3View_StartRendering(view);
do {
   Q3Geometry_Submit(curveObject, view);
} while (Q3View_EndRendering(view) == kQ3ViewStatusRetraverse);

// Dispose of the curve object.
Q3Object_Dispose(curveObject);

The equivalent drawing operation in immediate mode uses exactly the same code up to the point where the object is created. Instead of creating the retained object, we simply pass the TQ3NURBCurveData structure directly to the QuickDraw 3D function Q3NURBCurve_Submit to be rendered immediately:

// Render the curve directly.
Q3View_StartRendering(view);
do {
   Q3NURBCurve_Submit(&curveData, view);
} while (Q3View_EndRendering(view) == kQ3ViewStatusRetraverse);

CONTROLLING SUBDIVISION

QuickDraw 3D doesn't render NURB curves directly -- as it does, say, lines or triangles. To draw a NURB curve, the renderer has to break it up into a sequence of lines or polylines. The more lines it's broken up into, the smoother it looks, but of course the longer it takes to render. Before rendering a curve, you have to tell the renderer how finely you want it subdivided.

There are three ways of doing this, denoted by the values of an enumerated data type:

typedef enum TQ3SubdivisionMethod {
   kQ3SubdivisionMethodConstant,
   kQ3SubdivisionMethodWorldSpace,
   kQ3SubdivisionMethodScreenSpace
} TQ3SubdivisionMethod;

• The first method, kQ3SubdivisionMethodConstant, says to subdivide the curve into a polyline with a specified number of segments between each pair of joints.
• The second method, kQ3SubdivisionMethodWorldSpace, says to subdivide the curve so that the length of each line segment is no longer than a specified value, measured in world space.
• The third method, kQ3SubdivisionMethodScreenSpace, is similar to the second, but the measurement is done in screen space.

The following data structure specifies the subdivision method to use and the relevant parameter values:

typedef struct TQ3SubdivisionStyleData {
   TQ3SubdivisionMethod      method;
   float                     c1;
   float                     c2;
} TQ3SubdivisionStyleData;

NURB curves use only the c1 component; the other is for NURB surfaces. A couple of things to note:

• You should set both c1 and c2 to legitimate values. QuickDraw 3D doesn't know whether a curve or a surface is coming up, so it always checks both parameters for validity. If you're only drawing a curve, you may as well set c2 to the same value as c1.
• If you specify an unreasonable value for either parameter, QuickDraw 3D will substitute a more reasonable one and issue a warning. It won't let you subdivide a curve at a million positions!
• For method kQ3SubdivisionMethodConstant, c1 should be a whole number greater than 0; fractional values will be truncated.
• If you don't specify a subdivision style, the default value will be used.

Expanding on our example of immediate mode rendering, the following code will render our NURB curve with a five-segment polyline between each pair of knots:

TQ3SubdivisionStyleData         subdivData;
...
subdivData.method = kQ3SubdivisionMethodConstant;
subdivData.c1 = subdivData.c2 = 5;
...
Q3View_StartRendering(view);
do {
   Q3SubdivisionStyle_Submit(&subdivData, view);
   Q3NURBCurve_Submit(&curveData, view);
} while (Q3View_EndRendering(view) == kQ3ViewStatusRetraverse);

EDITING NURB CURVES

If you're rendering your curve in immediate mode, you can edit the curve by simply modifying its control points, weights, and knot vectors directly in the TQ3NURBCurveData structure. If you're using retained mode, QuickDraw 3D provides calls to retrieve and set individual control points and knots:

TQ3Status Q3NURBCurve_GetControlPoint(TQ3GeometryObject curve,
      unsigned long pointIndex, TQ3RationalPoint4D *point4D);

TQ3Status Q3NURBCurve_SetControlPoint(TQ3GeometryObject curve,
      unsigned long pointIndex, const TQ3RationalPoint4D *point4D);
   
TQ3Status Q3NURBCurve_GetKnot(TQ3GeometryObject curve,
      unsigned long knotIndex, float *knotValue);

TQ3Status Q3NURBCurve_SetKnot(TQ3GeometryObject curve,
      unsigned long knotIndex, float knotValue);

Because we're not interacting with items that are objects themselves, there are no reference counts involved and no need to dispose of any data structures. Note, however, that if you edit a knot, the resulting knot vector must remain nondecreasing and follow the limitations described earlier for multiple knots.

You may have noticed that there are no calls to add, delete, or reorder control points or knots. Instead, QuickDraw 3D provides calls for retrieving and replacing the entire TQ3NURBCurveData structure from the retained object:

TQ3Status Q3NURBCurve_GetData(TQ3GeometryObject curve,
             TQ3NURBCurveData *nurbCurveData);

TQ3Status Q3NURBCurve_SetData(TQ3GeometryObject curve, 
             const TQ3NURBCurveData *nurbCurveData);

If you want to change the number of control points and knots in a curve, you have to make a local copy of the data structure you obtain from Q3NURBCurve_GetData (making sure to allocate extra space for the new knots and control points), modify the arrays in the local copy, and store it back into the object with Q3NURBCurve_SetData. You must then call the following routine to dispose of the data you received from Q3NURBCurve_GetData:

TQ3Status Q3NURBCurve_EmptyData(TQ3NURBCurveData *nurbCurveData);

However, if you're going to be modifying the NURB curve frequently, you should probably be working in immediate mode and not using a retained object at all.

KNOT INSERTION

In general, the more control points we define for a NURB curve, the more control we have over its shape. It would seem reasonable that we could add more control points without changing the shape of the curve, and in fact this turns out to be true. Remember, though, that there's a fundamental relationship among the knots, the control points, and the order of the curve: the number of knots is equal to the number of control points plus the order. For example, a cubic curve (order 4) with 9 control points will require 13 knots. So every time we add a control point, we also have to add an extra knot -- and make sure all the control points are in the correct locations to keep the curve's shape the same as before.

In practice, we actually take the reverse approach: we decide where to add a new knot, then compute the location of the corresponding new control point (as well as the new locations of some of the existing ones). For example, if we take the curve depicted earlier in Figure 9 and insert a new knot at t = 3.6, we get a new curve with exactly the same shape but with a new set of control points (Figure 22).

Figure 22. Inserting a knot

This operation of knot insertion is a fundamental one in working with NURB curves. It's directly useful in both modifying (editing) and rendering curves, and can also be used to convert a NURB curve to Bézier representation. After a brief discussion of the mathematical algorithm for inserting a knot, we'll look at some example C code for implementing it.

THE ALGORITHM

We start with a NURB curve represented by

with a knot vector {x0, x1, . . . , xn+k-1}. Suppose we want to add a new knot xnew, where xi < xnew <= xi+1. The new knot vector ^x is simply the old knot vector with xnew inserted between xi and xi+1. The new curve will be defined by

with knot vector ^x.

Now we have to figure out not only where the new control point is located and where it goes in the ordered vector of control points, but also how to adjust some of the existing control points to keep the shape of the curve unchanged; this process yields the new control point vector, ^B. It turns out that the relationship between the old and new control points is

where [[alpha]] is defined by

The proof of this is relatively simple, but we don't have the time or space to go into it here. For a full discussion, see Curves and Surfaces in Computer Aided Geometric Design.

THE IMPLEMENTATION

Listing 2 shows a function to implement this basic algorithm. The function, which is included on this issue's CD, accepts a QuickDraw 3D NURB-curve data structure as an argument, along with the value of the new knot to insert, and returns a new data structure representing the same curve with the new knot inserted. For brevity, the function performs no range checking on the inserted knot, but simply assumes that it falls within the legal range and that the resulting knot vector obeys the usual limitations on multiple knots. Note also that the code shown here does no checking on the results of memory allocation requests, though of course you should always perform such checks in real life.

Listing 2. Inserting a knot

static TQ3NURBCurveData *InsertKnot
   (TQ3NURBCurveData    *oldCurveData, // Old curve
    float              tNew)           // Knot to insert

{
   TQ3NURBCurveData    *newCurveData;  // New curve after adding knot
   unsigned long       k;              // Order of curve
   unsigned long       n;              // Number of control points
   TQ3RationalPoint4D  *b;             // Old control point vector
   TQ3RationalPoint4D  *bHat;          // New control point vector
   float               *x;             // Old knot vector
   float               *xHat;          // New knot vector
   float               alpha;          // Interpolation ratio
   unsigned long       i;              // Knot to insert after
   unsigned long       j;              // Knot index for search
   TQ3Boolean          foundIndex;     // Insertion index found?
   
   // Set up local variables for readability.
   k = oldCurveData->order;
   n = oldCurveData->numPoints;
   x = oldCurveData->knots;
   b = oldCurveData->controlPoints;
   
   // Allocate space for new control points and knot vector.
   bHat = malloc((n + 1) * sizeof(TQ3RationalPoint4D));
   xHat = malloc((n + k + 1) * sizeof(float));
   
   // Allocate data structure for new curve.
   newCurveData = malloc(sizeof(TQ3NURBCurveData));
   newCurveData->order = k;
   newCurveData->numPoints = n + 1;
   newCurveData->controlPoints = bHat; 
   newCurveData->knots = xHat;
   newCurveData->curveAttributeSet = 
      (oldCurveData->curveAttributeSet == NULL)
          ? NULL
          : Q3Object_Duplicate(oldCurveData->curveAttributeSet);
   
   // Find where to insert the new knot. 
   for (j = 0, foundIndex = kQ3False; j < n + k; j++) {
      if (tNew > x[j] && tNew <= x[j + 1]) {
         i = j;
         foundIndex = kQ3True;
         break;
      }
   }
   
   // Return if not found.
   if (!foundIndex) {
      return (NULL);
   }
   // Copy knots to new vector.
   for (j = 0; j < n + k + 1; j++) {
      if (j <= i) {
         xHat[j] = x[j];
      } else if (j == i + 1) {
         xHat[j] = tNew;
      } else {
         xHat[j] = x[j - 1];
      }
   }

   // Compute position of new control point and new positions of
   // existing ones.
   for (j = 0; j < n + 1; j++) {
      if (j <= i - k + 1) {
         alpha = 1;
      } else if (i - k + 2 <= j && j <= i) {
         if (x[j + k - 1] - x[j] == 0) {
            alpha = 0;
         } else {
            alpha = (tNew - x[j]) / (x[j + k - 1] - x[j]); 
         }
      } else {
         alpha = 0;
      }
      
      if (alpha == 0) {
         bHat[j] = b[j - 1];
      } else if (alpha == 1) {
         bHat[j] = b[j];
      } else {
         bHat[j].x = (1 - alpha) * b[j - 1].x + alpha * b[j].x;
         bHat[j].y = (1 - alpha) * b[j - 1].y + alpha * b[j].y;
         bHat[j].z = (1 - alpha) * b[j - 1].z + alpha * b[j].z;
         bHat[j].w = (1 - alpha) * b[j - 1].w + alpha * b[j].w;
      }
   }

   return (newCurveData);
}

EVALUATING NURB CURVES

Recall from our earlier discussion that if we have two knots at the same location, we lose one degree of continuity; with three identical knots, we lose two degrees of continuity; and so on. This process can be repeated until, when we reach k-1 identical knots (where k is the order of the curve), we have no continuity at all at the given point. In this case, the curve at that point coincides directly with a control point, as we saw in Figure 15.

We've just seen that we can add a knot xnew and calculate the new control points. If we take this "new" curve (really just the old one with more knots) and add in that same knot again and again, until we have k-1 knots in the same place, we'll end up with a control point that lies exactly at ^ Q(xnew). We can use this technique to calculate the location of a particular point on the NURB curve: simply keep inserting knots at the point of interest until there are k-1 of them, at which time the newest control point created will lie at the desired point on the curve.

We can also use this approach to render a curve: by adding enough knots at some number of successive points in time t, we'll end up with a list of evaluated points on the curve, which we can then render as a polyline. The greater the number of evaluation points, the more segments the polyline will have, and the more closely the resulting image will approximate the curve.

This isn't the most efficient algorithm; a number of better alternatives are available. For example, see An Introduction to Splines for Use in Computer Graphics and Geometric Modeling for a description of the Oslo algorithm, which is significantly more efficient if you're adding more than a few knots.*

NURB CURVES AND BÉZIER CURVES

If you're familiar with Bézier curves, you may be wondering how they relate to NURB curves. In particular, if your application currently uses Bézier curves, how can you draw them when QuickDraw 3D currently only supports NURB curves? Although a thorough treatment of the subject is beyond the scope of this article, you'll be happy to learn that Bézier curves can actually be viewed as a subset of NURB curves. As a result, converting from Bézier to NURB representation turns out to be trivial.

CONVERTING BÉZIER TO NURB CURVES

Here's all it takes to convert a Bézier curve to NURB format:

• Use the Bézier control points as the NURB control points. If the Bézier control points are rational (that is, if they have four components {x, y, z, w}), make sure they're in homogeneous rather than weighted Euclidean form. If they're nonrational (have no w component), simply set w = 1.0 for each NURB control point.
• Set the order of the NURB curve to the number of control points. Bézier curves typically have three or four control points, corresponding to quadratic (order-3) or cubic (order-4) NURB curves, respectively.
• Create a knot vector with 2k elements, where k is the order of the curve. Set the first k knots to 0.0 and the last k to 1.0.

Listing 3 shows a function to perform the conversion (it's included on this issue's CD). The Bézier curve is assumed to be represented by a data structure of the form

typedef struct BezierCurve {
   unsigned int   order;
   Point3D         *controlPoints;
} BezierCurve;

Listing 3. Converting a Bézier curve to NURB format

TQ3NURBCurveData *BezierToNURBCurve(BezierCurve *bezCurve)
{
   TQ3NURBCurveData  *nurbCurveData; // NURB curve data structure
   unsigned long     k;              // Order of curve
   Point3D           *b;             // Bezier control point vector
   unsigned long     i;              // Control point or knot index
   
   // Set up local variables for readability.
   k = bezCurve->order;
   b = bezCurve->controlPoints;
   
   // Allocate data structure for new curve.
   nurbCurveData = malloc(sizeof(TQ3NURBCurveData));
   nurbCurveData->order = k;
   nurbCurveData->numPoints = k;
   nurbCurveData->controlPoints =
      malloc(k*sizeof(TQ3RationalPoint4D));
   nurbCurveData->knots = malloc(2*k*sizeof(float);
   
   // Create the control points.
   for (i = 0; i < k; i++) {
      TQ3RationalPoint4D_Set(&nurbCurveData->controlPoints[i],
                              b[i].x, b[i].y, b[i].z, 1.0);
   }
   
   // Create the knots.
   for (i = 0; i < k; i++) {
      nurbCurveData->knots[i] = 0.0;
      nurbCurveData->knots[i + k] = 1.0;
   }
   
   // Set attributes here, if desired.
   nurbCurveData->nurbCurveAttributes = NULL;
   
   return (nurbCurveData);
}

where the number of control points is equal to the order of the curve. The function returns a TQ3NURBCurveData structure representing the equivalent NURB curve. Once again, we've saved code space by leaving out the necessary checks on the results of memory allocation requests.

CONVERTING NURB TO BÉZIER CURVES

Converting a NURB curve to Bézier format is more complicated than the other way around. As we've just seen, any Bézier curve can be represented by a particular type of NURB curve, having half its knots at one end and half at the other. The converse, however, isn't true: an arbitrary NURB curve can't, in general, be represented by a single Bézier curve. In fact, it generally requires several Béziers to represent a single NURB curve: one for each distinct segment of the curve, as defined by its knot vector.

Recall that each segment of a NURB curve is affected by some subset of the control points. If we take each segment and add knots to both ends, generating a new set of control points each time, until each end has a number of knots equal to the order of the curve, the result will be a Bézier representation of that particular segment. Do this for each segment, and we'll end up with a series of Bézier curves that, taken together, look exactly like the original NURB curve.

DESIGNING WITH NURB CURVES

The topic of how to use NURB curves in design could easily fill a book; we'll have to be content with just a brief discussion, along with some pointers for further reading.

The most obvious capabilities an application program can offer for creating and modifying NURB curves are

• interactive placement and movement of control points
• interactive placement and movement of knots
• interactive setting and modification of control-point weights

These capabilities can be moderately effective, but actually using them to model a desired shape turns out to be difficult and awkward. In addition, modifying a control point, knot, or weight will generally affect parts of the curve that the user wants to remain unchanged.

One problem that has been explored extensively is that of automatically creating a curve that goes through (interpolates) a given set of points, which may have been interactively placed by the user or perhaps obtained by some sort of data sampling. Indeed, it might be said that this was one of the original motivations for the mathematical development of spline curves. The first straightforward attempts yielded less than satisfactory results, but later efforts weren't too bad and may be useful if the curve must pass exactly through the given points. Often, however, we only need to approximate the given set of points with a spline curve. The points may have been obtained by sampling the user's freehand drawing with a mouse or tablet, or perhaps by measuring a physical object or extracting edge information from a glyph in a bitmapped font. In these cases, we probably want to preserve features such as endpoints and corners, but the remaining data samples may be noisy or nonsmooth and need not be fitted exactly. Techniques for both exact and approximate fitting can be found in Phoenix: An Interactive Curve Design System Based on the Automatic Fitting of Hand-Sketched Curves and A User Interface Model and Tools for Geometric Design. These techniques can also be adapted for use in modifying an existing curve, whether it was generated in the usual way or via one of these fitting algorithms.

CURVING ON

Well, there you have it: more than you probably wanted to know about NURB curves, plus some free code to boot. Look for a possible upcoming article on NURB surfaces, and how NURB curves and surfaces can be used together for designing objects and controlling motion.

BIBLIOGRAPHY AND RECOMMENDED READING

This article only scratches the surface of the theory underlying NURB curves. The following is a list of books and articles referred to in this article, as well as others you may want to investigate for further information.

• Curves and Surfaces for Computer Aided Geometric Design: A Practical Guide by Gerald Farin (Academic Press, 1990).
• Curves and Surfaces in Computer Aided Geometric Design by Fujio Yamaguchi (Springer-Verlag, 1988).
• Fundamentals of Computer Aided Geometric Design by Josef Hoschek and Dieter Lasser (A. K. Peters, 1993).
• An Introduction to Splines for Use in Computer Graphics and Geometric Modeling by Richard H. Bartels, John C. Beatty, and Brian A. Barsky (Morgan Kaufman Publishers, 1987).
• Mathematical Elements for Computer Graphics by David F. Rogers and J. Alan Adams (McGraw-Hill, 1976).
• NURB Curves and Surfaces from Projective Geometry to Practical Use by Gerald Farin (A. K. Peters, 1995).
• Phoenix: An Interactive Curve Design System Based on the Automatic Fitting of Hand-Sketched Curves by Philip J. Schneider (master's thesis, University of Washington, 1988).
• "A Survey of Curve and Surface Methods in CAGD" by Wolfgang Böhm, Gerald Farin, and Jürgen Kahman, in Computer Aided Geometric Design, volume 1 (1984).
• A User Interface Model and Tools for Geometric Design by Michael J. Banks (master's thesis, University of Utah, 1989).

PHILIP J. SCHNEIDER (pjs@apple.com) is the longest-surviving member of the QuickDraw 3D team. He lives with his wife Suzanne and son Dakota out in the middle of a redwood forest in the Santa Cruz mountains, pretending he does so because "it's more affordable." People who are taken in by that malarkey probably also believe he doesn't like driving a two-lane country highway to work every day, and would rather be stuck in traffic jams on the interstate freeway with flatlanders. His current projects include trying to single-handedly bring up the worldwide level of computer technology to what he finds in the science fiction novels he reads voraciously, and teaching his 18-month-old son to change his own wet diapers in the middle of the night.*

Thanks to our technical reviewers Pablo Fernicola, Jim Mildrew, Klaus Strelau, and Nick Thompson.*