100% found this document useful (1 vote)

2K views

Computer Graphics With Using Opengl

Computer graphics are pictures generated by a computer, including those in books, magazines, movies, and more. The field involves both the tools used to create graphics (hardware and software) and the study of how to programmatically generate pictures. Some key elements that make up computer-generated pictures include polylines (connected straight lines), text, filled regions, and raster images. Attributes like color and thickness also impact how graphic primitives appear.

Uploaded by

Suhas Maringanti

Available Formats

Download as PDF, TXT or read online on Scribd

100% found this document useful (1 vote)

2K views

Computer Graphics With Using Opengl

Uploaded by

Suhas Maringanti

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 385

(excerpt For ECE660, Fall 1999) - F.S.Hill,Jr.

CHAPTER 1. Introduction to Computer Graphics

Begin at the beginning, the King said gravely,
and go on till you come to the end; then stop.
Lewis Carroll, Alice in Wonderland
The machine does not isolate
the man from the great
problems of nature
but plunges him more deeply
into them
Antoine de Saint-Exupry

Any sufficiently advanced

technology is indistinguishable from magic.
Arthur C. Clarke

1.1 What is computer graphics?

Good question. People use the term computer graphics to mean different things in different contexts.
Most simply, computer graphics are pictures that are generated by a computer. Everywhere you look
today there are examples to be found, especially in magazines and on television. This book was typeset
using a computer: every character (even this one: G) was drawn from a library of character shapes
stored in computer memory. Books and magazines abound with pictures created on a computer. Some
look so natural you cant distinguish them from photographs of a real scene. Others have an artificial or
surreal feeling, intentionally fashioned to achieve some visual effect. And movies today often show scenes
that never existed, but were carefully crafted by computer, mixing the real and the imagined.
Computer graphics also refers to the tools used to make such pictures. The purpose of this book is to
show what the tools are and how to apply them. There are both hardware and software tools. Hardware
tools include video monitors and printers that display graphics, as well as input devices like a mouse or
trackball that let a user point to items and draw figures. The computer itself, of course, is a hardware tool,
along with its special circuitry to facilitate graphical display or image capture.
As for software tools, you are already familiar with the usual ones: the operating system, editor, compiler, and
debugger, that are found in any programming environment. For graphics there must also be a collection of
graphics routines that produce the pictures themselves. For example, all graphics libraries have functions to
draw a simple line or circle (or characters such as G). Some go well beyond this, containing functions to
draw and manage windows with pull-down menus and dialog boxes, or to set up a camera in a threedimensional coordinate system and to make snapshots of objects stored in some data base.
In this book we show how to write programs that utilize graphics libraries, and how to add functionality to
them. Not too long ago, programmers were compelled to use highly device dependent libraries,
designed for use on one specific computer system with one specific display device type. This made it very
difficult to port a program to another system, or to use it with another device: usually the programmer
had to make substantial changes to the program to get it to work, and the process was time-consuming
and highly error-prone. Happily the situation is far better today. Device independent graphics libraries are
now available that allow the programmer to use a common set of functions within an application, and to
run the same application on a variety of systems and displays. OpenGL is such a library, and serves as the
main tool we use in this book. The OpenGL way of creating graphics is used widely in both universities
and industry. We begin a detailed discussion of it in Chapter 2.

Hill - Chapter 1

9/9/99

page 1

Finally, computer graphics often means the whole field of study that involves these tools and the
pictures they produce. (So its also used in the singular form: computer graphics is...). The field is often
acknowledged to have started in the early 1960s with Ivan Sutherlands pioneering doctoral thesis at MIT
on Sketchpad [ ref]. Interest in graphics grew quickly, both in academia and industry, and there were
rapid advances in display technology and in the algorithms used to manage pictorial information. The
special interest group in graphics, SIGGRAPH1, was formed in 1969, and is very active today around the
world. (The must-not-miss annual SIGGRAPH meeting now attracts 30,000 participants a year.) More
can be found at http://www.siggraph.org. Today there are hundreds of companies around the world having
some aspect of computer graphics as their main source of revenue, and the subject of computer graphics is
taught in most computer science or electrical engineering departments.
Computer graphics is a very appealing field of study. You learn to write programs that create pictures,
rather than streams of text or numbers. Humans respond readily to pictorial information, and are able to
absorb much more information from pictures than from a collection of numbers. Our eye-brain systems
are highly attuned to recognizing visual patterns. Reading text is of course one form of pattern
recognition: we instantly recognize character shapes, form them into words, and interpret their meaning.
But we are even more acute when glancing at a picture. What might be an inscrutable blather of numbers
when presented as text becomes an instantly recognizable shape or pattern when presented graphically.
The amount of information in a picture can be enormous. We not only recognize whats in it, but also
glean a world of information from its subtle details and texture
People study computer graphics for many reasons. Some just want a better set of tools for plotting curves
and presenting the data they encounter in their other studies or work. Some want to write computeranimated games, while others are looking for a new medium for artistic expression. Everyone wants to be
more productive, and to communicate ideas better, and computer graphics can be a great help.
There is also the input side. A program generates output pictures or otherwise from a combination
of the algorithms executed in the program and the data the user inputs to the program. Some programs
accept input crudely through characters and numbers typed at the keyboard. Graphics program, on the
other hand, emphasize more familiar types of input: the movement of a mouse on a desktop, the strokes of
a pen on a drawing tablet, or the motion of the users head and hands in a virtual reality setting. We
examine many techniques of interactive computer graphics in this book; that is, we combine the
techniques of natural user input with those that produce pictorial output.
(Section 1.2 on uses of Computer Graphics deleted.)

1.3. Elements of Pictures Created in Computer Graphics.

What makes up a computer drawn picture? The basic objects out of which such pictures are composed are
called output primitives. One useful categorization of these is:
polylines
text
filled regions
raster images
We will see that these types overlap somewhat, but this terminology provides a good starting point. We
describe each type of primitive in turn, and hint at typical software routines that are used to draw it. More
detail on these tools is given in later chapters, of course. We also discuss the various attributes of each
output primitive. The attributes of a graphic primitive are the characteristics that affect how it appears,
such as color and thickness.

SIGGRAPH is a Special Interest Group in the ACM: the Association for Computing Machinery.

Hill - Chapter 1

9/9/99

page 2

1.3.1. Polylines.
A polyline is a connected sequence of straight lines. Each of the examples in Figure 1.8 contain several
polylines: a). one polyline extends from the nose of the dinosaur to its tail; the plot of the mathematical
function is a single polyline, and the wireframe picture of a chess pawn contains many polylines that
outline its shape .

Figure 1.8. a). a polyline drawing of a dinosaur (courtesy of Susan Verbeck),

b). a plot of a mathematical function,
c). a wireframe rendering of a 3D object.
Note that a polyline can appear as a smooth curve. Figure 1.9 shows a blow-up of a curve revealing its
underlying short line segments. The eye blends them into an apparently smooth curve.
blow-up

Figure 1.9. A curved line made up of straight line segments.

Pictures made up of polylines are sometimes called line drawings. Some devices, like a pen plotter, are
specifically designed to produce line drawings.
The simplest polyline is a single straight line segment. A line segment is specified by its two endpoints,
say (x1, y1) and (x2, y2). A drawing routine for a line might look like drawLine(x1, y1, x2, y2);
It draws a line between the two endpoints. We develop such a tool later, and show many examples of its
use. At that point we get specific about how coordinates like x1 are represented (by integers or by real
numbers), and how colors can be represented in a program.
A special case arises when a line segment shrinks to a single point, and is drawn as a dot. Even the
lowly dot has important uses in computer graphics, as we see later. A dot might be programmed using the
routine
drawDot(x1, y1);
When there are several lines in a polyline each one is called an edge, and two adjacent lines meet at a
vertex. The edges of a polyline can cross one another, as seen in the figures. Polylines are specified as a
list of vertices, each given by a coordinate pair:
(x0, y0), (x1, y1), (x2, y2), ...., (xn, yn)

Hill - Chapter 1

(1.1)

9/9/99

page 3

For instance, the polyline shown in Figure 1.10 is given by the sequence (2, 4), (2, 11), (6, 14), (12, 11),
(12, 4), .... (what are the remaining vertices in this polyline?).

Figure 1.10. An example polyline.

To draw polylines we will need a tool such as: drawPolyline( poly);
where the variable poly is a list containing all the endpoints (xi, yi) in some fashion. There are various
ways to capture a list in a program, each having its advantages and disadvantages.
A polyline need not form a closed figure, but if the first and last points are connected by an edge the
polyline is a polygon. If in addition no two edges cross, the polygon is called simple. Figure 1.11 shows
some interesting polygons; only A and D are simple. Polygons are fundamental in computer graphics,
partly because they are so easy to define, and many drawing (rendering) algorithms have been finely tuned
to operate optimally with polygons. Polygons are described in depth in Chapter 3.
A).
C).
B).

E).
D).

Figure 1.11. Examples of polygons..

Attributes of Lines and Polylines.
Important attributes of a polyline are the color and thickness of its edges, the manner in which the edges
are dashed, and the manner in which thick edges blend together at their endpoints. Typically all of the
edges of a polyline are given the same attributes.
The first two polylines in Figure 1.12 are distinguished by the line thickness attribute. The third polyline
is drawn using dashed segments.

Figure 1.12. Polylines with different attributes.

Hill - Chapter 1

9/9/99

page 4

When a line is thick its ends have shapes, and a user must decide how two adjacent edges join. Figure
1.13 shows various possibilities. Case a) shows butt-end lines that leave an unseemly crack at the
joint. Case b) shows rounded ends on the lines so they join smoothly, part c) shows a mitered joint, and
part d) shows a trimmed mitered joint. Software tools are available in some packages to allow the user to
choose the type of joining. Some methods are quite expensive computationally.
a).

b).

c).

d).

Figure 1.13. Some ways of joining two thick lines in a polyline.

The attributes of a polyline are sometimes set by calling routines such as setDash(dash7) or
setLineThickness(thickness).

1.3.2. Text.
Some graphics devices have two distinct display modes, a text mode and a graphics mode. The text mode
is used for simple input/output of characters to control the operating system or edit the code in a program.
Text displayed in this mode uses a built-in character generator. The character generator is capable of
drawing alphabetic, numeric, and punctuation characters, and some selection of special symbols such as
, , and . Usually these characters cant be placed arbitrarily on the display but only in some row and
column of a built-in grid.
A graphics mode offers a richer set of character shapes, and characters can be placed arbitrarily. Figure
1.14 shows some examples of text drawn graphically.

Big Text
Little Text

Shadow Text

Rotated TextOutlined text

SMALLCAPS
Figure 1.14. Some text drawn graphically.
A tool to draw a character string might look like: drawString(x, y, string); It places the
starting point of the string at position (x, y), and draws the sequence of characters stored in the variable
string.
Text Attributes.
There are many text attributes, the most important of which are typeface, color, size, spacing, and
orientation.
Font. A font is a specific set of character shapes (a typeface) in a particular style and size. Figure 1.15
shows various character styles.

Hill - Chapter 1

9/9/99

page 5

Times
Times bold
Times italic

Helvetica
Helvetica bold
Helvetica italic

Courier
Courier bold
Courier italic
Figure 1.15. Some examples of fonts.
The shape of each character can be defined by a polyline (or more complicated curves such as Bezier
curves see Chapter 11), as shown in Figure 1.16a, or by an arrangement of dots, as shown in part b.
Graphics packages come with a set of predefined fonts, and additional fonts can be purchased from
companies that specialize in designing them.

B
Figure 1.16. A character shape defined by a polyline and by a pattern of dots.
Orientation of characters and strings: Characters may also be drawn tilted along some direction. Tilted
strings are often used to annotate parts of a graph. The graphic presentation of high-quality text is a
complex subject. Barely perceptible differences in detail can change pleasing text into ugly text. Indeed,
we see so much printed material in our daily lives that we subliminally expect characters to be displayed
with certain shapes, spacings, and subtle balances.
1.3.3. Filled Regions
The filled region (sometimes called fill area) primitive is a shape filled with some color or pattern. The
boundary of a filled region is often a polygon (although more complex regions are considered in Chapter
4). Figure 1.17 shows several filled polygons. Polygon A is filled with its edges visible, whereas B is filled
with its border left undrawn. Polygons C and D are non-simple. Polygon D even contains polygonal holes.
Such shapes can still be filled, but one must specify exactly what is meant by a polygons interior, since
filling algorithms differ depending on the definition. Algorithms for performing the filling action are
discussed in Chapter 10.
B

Hill - Chapter 1

9/9/99

page 6

Figure 1.17. Examples of filled Polygons.

To draw a filled polygon one would use a routine like: fillPolygon(poly, pattern);
where the variable poly holds the data for the polygon - the same kind of list as for a polyline - and the
variable pattern is some description of the pattern to be used for filling. We discuss details for this in
Chapter 4.
Figure 1.18 shows the use of filled regions to shade the different faces of a 3D object. Each polygonal
face of the object is filled with a certain shade of gray that corresponds to the amount of light that would
reflect off that face. This makes the object appear to be bathed in light from a certain direction. Shading of
3D objects is discussed in Chapter 8.

Figure 1.18. Filling polygonal faces of 3D objects to suggest proper shading.

The attributes of a filled region include the attributes of the enclosing border, as well as the pattern and
color of the filling.
1.3.4. Raster Image.
Figure 1.19a shows a raster image of a chess piece. It is made up of many small cells, in different
shades of gray, as revealed in the blow-up shown in Figure 1.19b. The individual cells are often called
pixels (short for picture elements). Normally your eye cant see the individual cells; it blends them
together and synthesizes an overall picture.

Figure 1.19. a). A raster image of a chess piece. b). A blow-up of the image. (Raytracing courtesy of
Andrew Slater)

Hill - Chapter 1

9/9/99

page 7

A raster image is stored in a computer as an array of numerical values. This array is thought of as being
rectangular, with a certain number of rows and a certain number of columns. Each numerical value
represents the value of the pixel stored there. The array as a whole is often called a pixel map. The term
bitmap is also used, (although some people think this term should be reserved for pixel maps wherein
each pixel is represented by a single bit, having the value 0 or 1.)
Figure 1.20 shows a simple example where a figure is represented by a 17 by 19 array (17 rows by 19
columns) of cells in three shades of gray. Suppose the three gray levels are encoded as the values 1, 2, and
7. Figure 1.20b shows the numerical values of the pixel map for the upper left 6 by 8 portion of the image.
a).
b).

2
2
2
2
2
2

2
2
2
2
7
7

2
2
2
7
1
1

2
2
7
1
1
1

2
2
7
1
1
7

2
7
1
1
1
7

Figure 1.20. A simple figure represented as a bitmap.

How are raster images created? The three principal sources are:
1). Hand designed images.
A designer figures out what values are needed for each cell, and types them into memory. Sometimes a
paint program can be used to help automate this: the designer can draw and manipulate various
graphical shapes, viewing what has been made so far. When satisfied, the designer stores the result in
a file. The icon above was created this way.
2). Computed Images.
An algorithm is used to render a scene, which might be modeled abstractly in computer memory. As a
simple example, a scene might consist of a single yellow smooth sphere illuminated by a light source that
emanates orange light. The model contains descriptions of the size and position of the sphere, the
placement of the light source, and a description of the hypothetical camera that is to take the picture.
The raster image plays the role of the film in the camera. In order to create the raster image, an
algorithm must calculate the color of light that falls on each pixel of the image in the camera. This is the
way in which ray traced images such as the chess piece in Figure 1.20 are created; see Chapter 16.
Raster images also frequently contain images of straight lines. A line is created in an image by setting
the proper pixels to the line's color. But it can require quite a bit of computation to determine the
sequence of pixels that best fit the ideal line between two given end points. Bresenham's algorithm
(see Chapter 2) provides a very efficient approach to determining these pixels.
Figure 1.21a shows a raster image featuring several straight lines, a circular arc, and some text
characters. Figure 1.21b shows a close-up of the raster image in order to expose the individual pixels
that are on the lines. For a horizontal or vertical line the black square pixels line up nicely forming
a sharp line. But for the other lines and the arc the best collection of pixels produces only an
approximation to the true line desired. In addition, the result shows the dread jaggies that have a
relentless presence in raster images.

Hill - Chapter 1

9/9/99

page 8

Figure 1.21. a). a collection of lines and text. b). Blow-up of part a, having jaggies.
3). Scanned images.
A photograph or television image can be digitized as described above. In effect a grid is placed over the
original image, and at each grid point the digitizer reads into memory the closest color in its repertoire.
The bitmap is then stored in a file for later use. The image of the kitten in Figure 1.22 was formed this way.

Figure 1.22. A scanned image.

Because raster images are simply arrays of numbers, they can be subsequently processed to good effect by
a computer. For instance, Figure 1.23 shows three successive enlargements of the kitten image above.
These are formed by pixel replication (discussed in detail in Chapter 10). Each pixel has been replicated
three times in each direction in part a; by six times in part b, and by 12 times in part c.

Hill - Chapter 1

9/9/99

page 9

Figure 1.23. Three successive blow-ups of the kitten image. a). three times enlargement, b). six times
enlargement.

Figure 1.23c. Twelve times enlargement.

As another example, one often needs to clean up a scanned image, for instance to remove specks of
noise or to reveal important details. Figure 1.24a shows the kitten image with gray levels altered to
increase the contrast and make details more evident, and Figure 1.24b shows the effect of edge
enhancement, achieved by a form of filtering the image.

Figure 1.24. Examples of image enhancement

Figure 1.25 shows two examples of editing an image to accomplish some visual effect. Part a shows the
kitten image embossed, and part b shows it distorted geometrically.

Hill - Chapter 1

9/9/99

page 10

Figure 1.25. Examples of altering an image for visual effect.

1.3.5. Representation of gray shades and color for Raster Images
An important aspect of a raster image is the manner in which the various colors or shades of gray are
represented in the bitmap. We briefly survey the most common methods here.

1.3.5.1. Gray-scale Raster Images.

If there are only two pixel values in a raster image it is called bi-level. Figure 1.26a shows a simple bilevel image, representing a familiar arrow-shaped cursor frequently seen on a computer screen. Its raster
consists of 16 rows of 8 pixels each. Figure 1.26b shows the bitmap of this image as an array of 1s and
0s. The image shown at the left associates black with a 1 and white with a 0, but this association might
just as easily be reversed. Since one bit of information is sufficient to distinguish two values, a bilevel
image is often referred to as a 1 bit per pixel image.
b).

a).

1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0

0
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0

0
0
1
1
1
1
1
1
1
1
0
0
0
0
0
0

0
0
0
1
1
1
1
1
1
1
0
0
0
0
0
0

0
0
0
0
1
1
1
1
1
1
1
1
0
0
0
0

0
0
0
0
0
1
1
1
0
0
1
1
1
1
0
0

0
0
0
0
0
0
1
1
0
0
0
0
1
1
0
0

0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0

Figure 1.26. A bilevel image of a cursor, and its bitmap.

When the pixels in a gray-scale image take on more than two values, each pixel requires more than a
single bit to represent it in memory. Gray-scale images are often classified in terms of their pixel depth,
the number of bits needed to represent their gray levels. Since an n-bit quantity has 2n possible values,
there can be 2n gray levels in an image with pixel depth n. The most common values are:

2 bits/pixel produce 4 gray levels

4 bits/pixel produce 16 gray levels
8 bits/pixel produce 256 gray levels

Hill - Chapter 1

9/9/99

page 11

Figure 1.27 shows 16 gray levels ranging from black to white. Each of the sixteen possible pixel values is
associated with a binary 4-tuple such as 0110 or 1110. Here 0000 represents black, 1111 denotes white,
and the other 14 values represent gray levels in between.

0000

0001 0010

1110

1111

brightness
black

white

Figure 1.27. Sixteen levels of gray.

Many gray scale images2 employ 256 gray levels, since this usually gives a scanned image acceptable
quality. Each pixel is represented by some 8-bit values such as 01101110. The pixel value usually
represents "brightness", where black is represented by 00000000, white by 11111111, and a medium gray
by 10000000. Figure 1.23 seen earlier uses 256 gray levels.
Effect of Pixel depth: Gray-scale Quantization.
Sometimes an image that initially uses 8 bits per pixel is altered so that fewer bits per pixel are used. This
might occur if a particular display device is incapable of displaying so many levels, or if the full image
takes up too much memory. Figures 1.28 through 1.30 show the effect on the kitten image if pixel values
are simply truncated to fewer bits. The loss in fidelity is hardly noticeable for the images in Figure 1.28,
which use 6 and 5 bits/pixel (providing 64 and 32 different shades of gray, respectively).

.
Figure 1.28. The image reduced to 6 bits/pixel and 5 bits/pixel.
But there is a significant loss in quality in the images of Figure 1.29. Part a shows the effect of truncating
each pixel value to 4 bits, so there are only 16 possible shades of gray. For example, pixel value 01110100
is replaced with 0111. In part b the eight possible levels of gray are clearly visible. Note that some areas of
the figure that show gradations of gray in the original now show a lake of uniform gray. This is often
called banding, since areas that should show a gradual shift in the gray level instead show a sequence of
uniform gray bands.

Thousands are available on the Internet, frequently as Gif, Jpeg, or Tiff images.

Hill - Chapter 1

9/9/99

page 12

Figure 1.29. The image reduced to 4 bits/pixel and to 3 bits/pixel.

Figure 1.30 shows the cases of 2 and 1 bits/pixel. In part a the four levels are clearly visible and there is a
great deal of banding. In part b there is only black and white and much of the original image information
has been lost. In Chapter 10 we show techniques such as dithering for improving the quality of an image
when two few bits are used for each pixel.

Figure 1.30. The image reduced to 2 bits/pixel and 1 bit/pixel.

1.3.5.2. Color Raster Images.
Color images are desirable because they match our daily experience more closely than do gray-scale
images. Color raster images have become more common in recent years as the cost of high quality color
displays has come down. The cost of scanners that digitize color photos has also become reasonable.
Each pixel in a color image has a color value, a numerical value that somehow represents a color. There
are a number of ways to associate numbers and colors (see Chapter 12 for a detailed discussion), but one
of the most common is to describe a color as a combination of amounts of red, green, and blue light. Each
pixel value is a 3-tuple, such as (23, 14, 51), that prescribes the intensities of the red, green, and blue
light components in that order.

Hill - Chapter 1

9/9/99

page 13

The number of bits used to represent the color of each pixel is often called its color depth. Each value in
the (red, green, blue) 3-tuple has a certain number of bits, and the color depth is the sum of these values.
A color depth of 3 allows one bit for each component. For instance the pixel value (0, 1, 1) means that the
red component is off, but both green and blue are on. In most displays the contributions from each
component are added together (see Chapter 12 for exceptions such as in printing), so (0,1,1) would
represent the addition of green and blue light, which is perceived as cyan. Since each component can be
on or off there are eight possible colors, as tabulated in Figure 1.31. As expected, equal amounts of red,
green, and blue, (1, 1, 1), produce white.
color value
displayed
0,0,0
black
0,0,1
blue
0,1,0
green
0,1,1
cyan
1,0,0
red
1,0,1
magenta
1,1,0
yellow
1,1,1
white
Figure 1.31. A common correspondence between color value and perceived color.
A color depth of 3 rarely offers enough precision for specifying the value of each component, so larger
color depths are used. Because a byte is such a natural quantity to manipulate on a computer, many
images have a color depth of eight. Each pixel then has one of 256 possible colors. A simple approach
allows 3 bits for each of the red and the green components, and 2 bits for the blue component. But more
commonly the association of each byte value to a particular color is more complicated, and uses a color
look-up table, as discussed in the next section.
The highest quality images, known as true color images, have a color depth of 24, and so use a byte for
each component. This seems to achieve as good color reproduction as the eye can perceive: more bits
dont improve an image. But such images require a great deal of memory: three bytes for every pixel. A
high quality image of 1080 by 1024 pixels requires over three million bytes!
Plates 19 through 21 show some color raster images having different color depths. Plate 19 shows a full
color image with a color depth of 24 bits. Plate 20 shows the degradation this image suffers when the
color depth is reduced to 8 by simply truncating the red and green components to 3 bits each, and the blue
component to 2 bits. Plate 21 also has a color depth of 8, so its pixels contain only 256 colors, but the 256
particular colors used have been carefully chosen for best reproduction. Methods to do this are discussed
in Chapter 12.
author-supplied
Plate 19. Image with 24 bits/pixel.
author-supplied
Plate 20. Image with 3 bits for red and green pixels, and two bits for blue pixels.
author-supplied
Plate 1.21. Image with 256 carefully chosen colors.

1.4. Graphics Display Devices

Hill - Chapter 1

9/9/99

page 20

expensive: ( frame buffer needs ~4 Mbyte )

24 bits/pixel

8
8
8

1.3 million pixels

inexpensive: (frame buffer needs ~ 1 Mbyte)
8 bits/pixel
24

8
8
1.3 million pixels

8
8

Figure 1.41. Comparison of two raster display systems.

System #1 (expensive): The first has a 24 bit/pixel frame buffer and no LUT, so each of its 1.3
million pixels can be set to any one of 224 colors. Eight bits drive each of the DACs. (The
number 8 and the slash through the line into each DAC indicate the presence of eight bit lines
feeding into the DAC.) The amount of memory required for the frame buffer here is 1024 x 1280
x 24 bits which is almost 4 megabytes.
System #2 (inexpensive): The second has an 8 bit/pixel frame buffer along with a LUT, and the
LUT is 24 bits wide. The system can display 224 different colors, but only 256 at a time. The
amount of memory required for the frame buffer here is 1024 x 1280 x 8 which is about 1
megabyte. (The LUT requires a trivial 768 bytes of memory.) If memory costs a significant
amount per megabyte, this system is much less expensive than system #1.
Putting a LUT in an inexpensive system attempts to compensate for the small number of different
pixel values possible. The LUT allows the programmer to create a full set of colors, even though
a given image can contain only a restricted set of them.
Displays with LUTs are still quite common today because memory costs are still high. The
situation is changing rapidly, however, as memory prices plummet. Many reasonably priced
personal computers today have 24 bit frame buffers.
Practice Exercises.
1.4.1. Why not always have a LUT? Since a LUT is inexpensive, and offers the advantage of
flexibility, why not have a LUT even in a system with 24 bits per pixel?
1.4.2. Configure your own system. For each of the systems described below:
a). draw the circuit diagram, similar to Figure 1.41;
b). Label the number of bits associated with the frame buffer, the DACs, and the LUT (if
present);
c). Calculate (in bytes) the amount of storage required for the frame buffer and the LUT (if
present):
i). b = 15, no LUT;
ii). b = 15, w = 24;
iii). b = 8, w = 18;
iv). b = 12, no LUT.

Hill - Chapter 1

9/9/99

page 21

1.5. Graphics Input Primitives and Devices.

Many input devices are available that let the user control a computer. Whereas typing a command might
be awkward, it is natural to point to a particular object displayed on the screen to make a choice for the
next action.
You can look at an input device in two ways: what it is, and what it does. Each device is physically some
piece of machinery like a mouse, keyboard, or trackball. It fits in the hand in a certain way and is natural
for the user to manipulate. It measures these manipulations and sends corresponding numerical
information back to the graphics program.
We first look at what input devices do, by examining the kinds of data each sends to the program. We
then look at a number of input devices in common use today.
1.5.1. Types of Input Graphics Primitives.
Each device transmits a particular kind of data (e.g. a number, a string of characters, or a position) to the
program. The different types of data are called input primitives. Two different physical devices may
transmit the same type of data, so logically they generate the same graphics primitives.
The important input primitives are:
String. The string device is the most familiar, producing a string of characters and thus modeling the
action of a keyboard. When an application requests a string, the program pauses while the user types it in
followed by a termination character. The program then resumes with the string stored in memory.
Choice. A choice device reports a selection from a fixed number of choices. The programmer's model is a
bank of buttons, or a set of buttons on a mouse.
Valuator. A valuator produces a real value between 0.0 and 1.0, which can be used to fix the length of a
line, the speed of an action, or perhaps the size of a picture. The model in the programmer's mind is a
knob that can be turned from 0 to 1 in smooth gradations.
Locator. A basic requirement in interactive graphics is to allow the user to point to a position on the
display. The locator input device performs this function, because it produces a coordinate pair (x, y). The
user manipulates an input device (usually a mouse) in order to position a visible cursor to some spot and
then triggers the choice. This returns to the application the values of x and y, along with the trigger value.
Pick. The pick input device is used to identify a portion of a picture for further processing. Some graphics
packages allow a picture to be defined in terms of segments, which are groups of related graphics. The
package provides tools to define segments and to give them identifying names. When using pick(), the
user points to a part of a picture with some physical input device, and the package figures out which
segment is being pointed to. pick() returns the name of the segment to the application, enabling the
user to erase, move, or otherwise manipulate the segment.
The graphics workstation is initialized when an application starts running: among other things each
logical input function is associated with one of the installed physical devices.
1.5.2. Types of Physical Input Devices.
We look at the other side of input devices: the physical machine that is connected to the personal
computer or workstation.
Keyboard. All workstations are equipped with a keyboard, which sends strings of characters to the
application upon request. Hence a keyboard is usually used to obtain a string device. Some keyboards
have cursor keys or function keys, which are often used to produce choice input primitives.

Hill - Chapter 1

9/9/99

page 22

Buttons. Sometimes a separate bank of buttons is installed on a workstation. The user presses one of the
buttons to perform a choice input function.
Mouse. The mouse is perhaps the most familiar input device of all, as it is easy and comfortable to
operate. As the user slides the mouse over the desktop, the mouse sends the changes in its position to the
workstation. Software within the workstation keeps track of the mouse's position and moves a graphics
cursor a small dot or cross on the screen accordingly. The mouse is most often used to perform a
locate or a pick function. There are usually some buttons on the mouse that the user can press to
trigger the action.
Tablet. Like a mouse, a tablet is used to generate locate or pick input primitives. A tablet provides
an area on which the user can slide a stylus. The tip of the stylus contains a microswitch. By pressing
down on the stylus the user can trigger the logical function.
The tablet is particularly handy for digitizing drawings: the user can tape a picture onto the tablet surface
and then move the stylus over it, pressing down to send each new point to the workstation. A menu area is
sometimes printed on the tablet surface, and the user Picks a menu item by pressing down the stylus
inside one of the menu item boxes. Suitable software associates each menu item box with the desired
function for the application that is running.
Space Ball and Data Glove. The Space Ball and Data Glove are relatively new input devices. Both are
designed to give a user explicit control over several variables at once, by performing hand and finger
motions. Sensors inside each device pick up subtle hand motions and translate them into Valuator values
that get passed back to the application. They are particularly suited to situations where the hand
movements themselves make sense in the context of the program, such as when the user is controlling a
virtual robot hand, and watching the effects of such motions simulated on the screen.
(Section 1.6 Summary - deleted.)

1.7. For Further Reading.

A number of books provide a good introduction to the field of computer graphics. Hearn and Baker
[hearn94] gives a leisurely and interesting overview of the field with lots of examples. Foley and Van
Dam [foley93] and David Rogers[rogers98] give additional technical detail on the many kinds of graphics
input and output devices. An excellent series of five books known as Graphics Gems [gems], first
published in 1990, brought together many new ideas and gems from graphics researchers and
practitioners around the world.
There are also a number of journals and magazines that give good insight into new techniques in
computer graphics. The most accessible is the IEEE Computer Graphics and Applications, which often
features survey articles on new areas of effort with graphics. The classic repositories of new results in
graphics are the annual Proceedings of SIGGRAPH [SIGGRAPH], and the ACM Transactions on
Graphics [TOGS]. Another more recent arrival is the Journal of Graphics Tools [jgt].

Hill - Chapter 1

9/9/99

page 23

(For Ece660 - Fall, 1999)

CHAP 2. Getting Started: Drawing Figures

Machines exist; let us then exploit them to
create beauty , a modern beauty, while we are
about it. For we live in the twentieth century.
Aldous Huxley
Goals of the Chapter
To get started writing programs that produce pictures.
To learn the basic ingredients found in every OpenGL program
To develop some elementary graphics tools for drawing lines, polylines, and polygons.
To develop tools that allow the user to control a program with the mouse and keyboard.
Preview
Section 2.1 discusses the basics of writing a program that makes simple drawings. The importance of deviceindependent programming is discussed, and the characteristics of windows-based and event-driven programs
are described. Section 2.2 introduces the use of OpenGL as the device-independent application programmer
interface (API) that is emphasized throughout the book and shows how to draw various graphics primitives.
Sample drawings, such as a picture of the Big Dipper, a drawing of the Sierpinski gasket, and a plot of a
mathematical function illustrate the use of OpenGL. Section 2.3 discusses how to make pictures based on
polylines and polygons, and begins the building of a personal library of graphics utilities. Section 2.4
describes interactive graphics programming, whereby the user can indicate positions on the screen with the
mouse or press keys on the keyboard to control the action of a program. The chapter ends with a number of
case studies which embellish ideas discussed earlier and that delve deeper into the main ideas of the chapter.

2.1. Getting started making pictures.

Like many disciplines, computer graphics is mastered most quickly by doing it: by writing and testing
programs that produce a variety of pictures. It is best to start with simple tasks. Once these are mastered you
can try variations, see what happens, and move towards drawing more complex scenes.
To get started you need an environment that lets you write and execute programs. For graphics this
environment must also include hardware to display pictures (usually a CRT display which we shall call the
screen), and a library of software tools that your programs can use to perform the actual drawing of
graphics primitives.
Every graphics program begins with some initializations; these establish the desired display mode, and set up
a coordinate system for specifying points, lines, etc. Figure 2.1 shows some of the different variations one
might encounter. In part a) the entire screen is used for drawing: the display is initialized by switching it into
graphics mode, and the coordinate system is established as shown. Coordinates x and y are measured in
pixels, with x increasing to the right and y increasing downward.

Figure 2.1. Some common varieties of display layouts.

Computer Graphics

Chap 2

09/14/99

1:01 PM

page 1

In part b) a more modern "window-based" system is shown. It can support a number of different rectangular
windows on the display screen at one time. Initialization involves creating and opening a new window
(which we shall call the screen window1) for graphics. Graphics commands use a coordinate system that is
attached to the window: usually x increases to the right and y increases downward2. Part c) shows a variation
where the initial coordinate system is right side up, with y increasing upward3.
Each system normally has some elementary drawing tools that help to get started. The most basic has a name
like setPixel(x, y, color): it sets the individual pixel at location (x, y) to the color specified by color.
It sometimes goes by different names, such as putPixel(), SetPixel(), or drawPoint(). Along with
setPixel() there is almost always a tool to draw a straight line, line(x1, y1, x2, y2), that draws a
line between (x1, y1) and (x2, y2). In other systems it might be called drawLine() or Line(). The
commands
line(100, 50, 150, 80);
line(150, 80, 0, 290);
would draw the pictures shown in each system in Figure 2.1. Other systems have no line() command, but
rather use moveto(x, y) and lineto(x, y). They stem from the analogy of a pen plotter, where the
pen has some current position. The notion is that moveto(x, y) moves the pen invisibly to location (x,
y), thereby setting the current position to (x, y); lineto(x, y) draws a line from the current position to (x,
y), then updates the current position to this (x, y). Each command moves the pen from its current position to a
new position. The new position then becomes the current position. The pictures in Figure 2.1 would be drawn
using the commands
moveto(100, 50);
lineto(150, 80);
lineto(0, 290);
For a particular system the energetic programmer can develop a whole toolkit of sophisticated functions that
utilize these elementary tools, thereby building up a powerful library of graphics routines. The final graphics
applications are then written making use of this personal library.
An obvious problem is that each graphics display uses different basic commands to drive it, and every
environment has a different collection of tools for producing the graphics primitives. This makes it difficult to
port a program from one environment to another (and sooner or later everyone is faced with reconstructing a
program in a new environment): the programmer must build the necessary tools on top of the new
environments library. This may require major alterations in the overall structure of a library or application,
and significant programmer effort.
2.1.1. Device Independent Programming, and OpenGL.
It is a boon when a uniform approach to writing graphics applications is made available, such that the same
program can be compiled and run on a variety of graphics environments, with the guarantee that it will
produce nearly identical graphical output on each display. This is known as device independent graphics
programming. OpenGL offers such a tool. Porting a graphics program only requires that you install the
appropriate OpenGL libraries on the new machine; the application itself requires no change: it calls the same
functions in this library with the same parameters, and the same graphical results are produced. The OpenGL
way of creating graphics has been adopted by a large number of industrial companies, and OpenGL libraries
exist for all of the important graphics environments4.
OpenGL is often called an application programming interface (API): the interface is a collection of
routines that the programmer can call, along with a model of how the routines work together to produce
graphics. The programmer sees only the interface, and is therefore shielded from having to cope with the
specific hardware or software idiosyncrasies on the resident graphics system.
1 The word "window" is overused in graphics: we shall take care to distinguish the various instances of the term.
2 Example systems are unix workstations using X Windows, an IBM pc running Windows 95 using the basic Windows

Application Programming Interface, and an Apple Macintosh using the built-in QuickDraw library.
3 An example is any window-based system using OpenGL.
4 Appendix 1 discusses how to obtain and get started with OpenGL in different environments.
Computer Graphics Chap 2
09/14/99
1:01 PM
page 2

OpenGL is at its most powerful when drawing images of complex three dimensional (3D) scenes, as we shall
see. It might be viewed as overkill for simple drawings of 2D objects. But it works well for 2D drawing, too,
and affords a unified approach to producing pictures. We start by using the simpler constructs in OpenGL,
capitalizing for simplicity on the many default states it provides. Later when we write programs to produce
elaborate 3D graphics we tap into OpenGLs more powerful features.
Although we will develop most of our graphics tools using the power of OpenGL, we will also look under
the hood and examine how the classical graphics algorithms work. It is important to see how such tools
might be implemented, even if for most applications you use the ready-made OpenGL versions. In special
circumstances you may wish to use an alternative algorithm for some task, or you may encounter a new
problem that OpenGL does not solve. You also may need to develop a graphics application that does not use
OpenGL at all.
2.1.2. Windows-based programming.
As described above, many modern graphics systems are windows-based, and manage the display of multiple
overlapping windows. The user can move the windows around the screen using the mouse, and can resize
them. Using OpenGL we will do our drawing in one of these windows, as we saw in Figure 2.1c.
Event-driven programming.
Another property of most windows-based programs is that they are event-driven. This means that the program
responds to various events, such as a mouse click, the press of a keyboard key, or the resizing of a screen
window. The system automatically manages an event queue, which receives messages that certain events have
occurred, and deals with them on a first-come first-served basis. The programmer organizes a program as a
collection of callback functions that are executed when events occur. A callback function is created for each
type of event that might occur. When the system removes an event from the queue it simply executes the
callback function associated with the type of that event. For programmers used to building programs with a
do this, then do this, structure some rethinking is required. The new structure is more like: do nothing
until an event occurs, then do the specified thing.
The method of associating a callback function with an event type is often quite system dependent. But
OpenGL comes with a Utility Toolkit (see Appendix 1), which provides tools to assist with event
management. For instance
glutMouseFunc(myMouse);

// register the mouse action function

registers the function myMouse() as the function to be executed when a mouse event occurs. The prefix
glut indicates it is part of the OpenGL Utility Toolkit. The programmer puts code in myMouse() to handle
all of the possible mouse actions of interest.
Figure 2.2 shows a skeleton of an example main() function for an event-driven program. We will base most
of our programs in this book on this skeleton. There are four principle types of events we will work with, and
a glut function is available for each:
void main()
{
initialize things5
create a screen window
glutDisplayFunc(myDisplay);
// register the redraw function
glutReshapeFunc(myReshape);
// register the reshape function
glutMouseFunc(myMouse);
// register the mouse action function
glutKeyboardFunc(myKeyboard); // register the keyboard action function
perhaps initialize other things
glutMainLoop();
// enter the unending main loop
}
all of the callback functions are defined here
Figure 2.2. A skeleton of an event-driven program using OpenGL.

5 Notes shown in italics in code fragments are pseudocode rather than actual program code. They suggest the actions
that real code substituted there should accomplish.
Computer Graphics Chap 2
09/14/99
1:01 PM
page 3

glutDisplayFunc(myDisplay); Whenever the system determines that a screen window should be

redrawn it issues a redraw event. This happens when the window is first opened, and when the window is
exposed by moving another window off of it. Here the function myDisplay() is registered as the callback
function for a redraw event.
glutReshapeFunc(myReshape); Screen windows can be reshaped by the user, usually by dragging a
corner of the window to a new position with the mouse. (Simply moving the window does not produce a
reshape event.) Here the function myReshape() is registered with the reshape event. As we shall see,
myReshape() is automatically passed arguments that report the new width and height of the reshaped
window.
glutMouseFunc(myMouse); When one of the mouse buttons is pressed or released a mouse event is
issued. Here myMouse() is registered as the function to be called when a mouse event occurs. myMouse() is
automatically passed arguments that describe the mouse location and the nature of the button action.
glutKeyboardFunc(myKeyboard); This registers the function myKeyboard() with the event of
pressing or releasing some key on the keyboard. myKeyboard() is automatically passed arguments that tell
which key was pressed. Conveniently, it is also passed data as to the location of the mouse at the time the
key was pressed.

If a particular program does not use mouse interaction, the corresponding callback function need not be
registered or written. Then mouse clicks have no effect in the program. The same is true for programs that have
no keyboard interaction.
The final function shown in Figure 2.2 is glutMainLoop(). When this is executed the program draws the
initial picture and enters an unending loop, in which it simply waits for events to occur. (A program is normally
terminated by clicking in the go away box that is attached to each window.)
2.1.3. Opening a Window for Drawing.
The first task is to open a screen window for drawing. This can be quite involved, and is system dependent.
Because OpenGL functions are device independent, they provide no support for window control on specific
systems. But the OpenGL Utility Toolkit introduced above does include functions to open a window on
whatever system you are using.
Figure 2.3 fleshes out the skeleton above to show the entire main() function for a program that will draw
graphics in a screen window. The first five function calls use the toolkit to open a window for drawing with
OpenGL. In your first graphics programs you can just copy these as is: later we will see what the various
arguments mean and how to substitute others for them to achieve certain effects. The first five functions
initialize and display the screen window in which our program will produce graphics. We give a brief
description of what each one does.
// appropriate #includes go here see Appendix 1
void main(int argc, char** argv)
{
glutInit(&argc, argv); // initialize the toolkit
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB); // set the display mode
glutInitWindowSize(640,480); // set window size
glutInitWindowPosition(100, 150); // set the window position on screen
glutCreateWindow("my first attempt"); // open the screen window
// register the callback functions
glutDisplayFunc(myDisplay);
glutReshapeFunc(myReshape);
glutMouseFunc(myMouse);
glutKeyboardFunc(myKeyboard);
myInit();
glutMainLoop();

// additional initializations as necessary

// go into a perpetual loop

}
Figure 2.3. Code using the OpenGL utility toolkit to open the initial window for drawing.

Computer Graphics

Chap 2

09/14/99

1:01 PM

page 4

glutInit(&argc, argv); This function initializes the toolkit. Its arguments are the standard ones
for passing command line information; we will make no use of them here.
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB); This function specifies how the display
should be initialized. The built-in constants GLUT_SINGLE and GLUT_RGB, which are ORd together,
indicate that a single display buffer should be allocated and that colors are specified using desired
amounts of red, green, and blue. (Later we will alter these arguments: for example, we will use double
buffering for smooth animation.)
glutInitWindowSize(640,480); This function specifies that the screen window should initially
be 640 pixels wide by 480 pixels high. When the program is running the user can resize this window as
desired.
glutInitWindowPosition(100, 150); This function specifies that the windows upper left
corner should be positioned on the screen 100 pixels from the left edge and 150 pixels down from the top.
When the program is running the user can move this window wherever desired.
glutCreateWindow("my first attempt"); This function actually opens and displays the
screen window, putting the title my first attempt in the title bar.

The remaining functions in main() register the callback functions as described earlier, perform any
initializations specific to the program at hand, and start the main event loop processing. The programmer
(you) must implement each of the callback functions as well as myInit().

2.2. Drawing Basic Graphics Primitives.

We want to develop programming techniques for drawing a large number of geometric shapes that make up
interesting pictures. The drawing commands will be placed in the callback function associated with a redraw
event, such as the myDisplay() function mentioned above.
We first must establish the coordinate system in which we will describe graphical objects, and prescribe
where they will appear in the screen window. Computer graphics programming, it seems, involves an ongoing
struggle with defining and managing different coordinate systems. So we start simply and work up to more
complex approaches.
We begin with an intuitive coordinate system. It is tied directly to the coordinate system of the screen window
(see Figure 2.1c), and measures distances in pixels. Our first example screen window, shown in Figure 2.4, is
640 pixels wide by 480 pixels high. The x-coordinate increases from 0 at the left edge to 639 at the right edge.
The y-coordinate increases from 0 at the bottom edge to 479 at the top edge. We establish this coordinate
system later, after examining some basic primitives.

Figure 2.4. The initial coordinate system for drawing.

OpenGL provides tools for drawing all of the output primitives described in Chapter 1. Most of them, such as
points, lines, polylines, and polygons, are defined by one of more vertices. To draw such objects in OpenGL
you pass it a list of vertices. The list occurs between the two OpenGL function calls glBegin() and
glEnd(). The argument of glBegin() determines which object is drawn. For instance, Figure 2.5 shows
three points drawn in a window 640 pixels wide and 480 pixels high. These dots are drawn using the
command sequence:
Computer Graphics

Chap 2

09/14/99

1:01 PM

page 5

Figure 2.5. Drawing three dots.

glBegin(GL_POINTS);
glVertex2i(100, 50);
glVertex2i(100, 130);
glVertex2i(150, 130);
glEnd();
The constant GL_POINTS is built-into OpenGL. To draw other primitives you replace GL_POINTS with
GL_LINES, GL_POLYGON, etc. Each of these will be introduced in turn.
As we shall see later, these commands send the vertex information down a graphics pipeline, in which they
go through several processing steps (look ahead to Figure ????). For present purposes just think of them as
being sent more or less directly to the coordinate system in the screen window.
Many functions in OpenGL like glVertex2i() have a number of variations. The variations distinguish the
number and type of arguments passed to the function. Figure 2.6 shows how the such function calls are
formatted.

glVertex2i()
gl
library

basic
command

number of
arguments

type of
argument

Figure 2.6. Format of OpenGL commands.

The prefix gl indicates a function from the OpenGL library (as opposed to glut for the utility toolkit).
It is followed by the basic command root, then the number of arguments being sent to the function (this
number will often be 3 and 4 in later contexts), and finally the type of argument, (i for an integer, f for a
floating point value, etc., as we describe below). When we wish to refer to the basic command without regard
to the specifics of its arguments we will use an asterisk, as in glVertex*().
To generate the same three dot picture above, for example, you could pass it floating point values instead of
integers, using:
glBegin(GL_POINTS);
glVertex2d(100.0, 50.0);
glVertex2d(100.0, 130.0);
glVertex2d(150.0, 130.0);
glEnd();
On OpenGL Data Types.
OpenGL works internally with specific data types: for instance, functions such as glVertex2i() expect
integers of a certain size (32 bits). It is well known that some systems treat the C or C++ data type int as a
16 bit quantity, whereas others treat it as a 32 bit quantity. There is no standard size for a float or double
either. To insure that OpenGL functions receive the proper data types it is wise to use the built-in type names
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 6

like GLint or GLfloat for OpenGL types. The OpenGL types are listed in Figure 2.7. Some of these types
will not be encountered until later in the book.
suffix
data type
typical C or C++ type
OpenGL type name
b
8-bit integer
signed char
GLbyte
s
16-bit integer
short
GLshort
i
32-bit integer
int or long
GLint, GLsizei
f
32-bit floating point
float
GLfloat, GLclampf
d
64-bit floating point
double
GLdouble,GLclampd
ub
8-bit unsigned number
unsigned char
GLubyte,GLboolean
us
16-bit unsigned number
unsigned short
GLushort
ui
32-bit unsigned number
unsigned int or unsigned long
GLuint,Glenum,GLbitfield
Figure 2.7. Command suffixes and argument data types.
As an example, a function using suffix i expects a 32-bit integer, but your system might translate int as a
16-bit integer. Therefore if you wished to encapsulate the OpenGL commands for drawing a dot in a generic
function such as drawDot() you might be tempted to use:

danger: passes ints

void drawDot(int x, int y)
{
// draw dot at integer point (x, y)
glBegin(GL_POINTS);
glVertex2i(x, y);
glEnd();
}
which passes ints to glVertex2i(). This will work on systems that use 32-bit ints, but might cause
trouble on those that use 16-bit ints. It is much safer to write drawDot() as in Figure 2.8, and to use
GLints in your programs. When you recompile your programs on a new system GLint, GLfloat, etc.
will be associated with the appropriate C++ types (in the OpenGL header GL.h see Appendix 1) for that
system, and these types will be used consistently throughout the program.
void drawDot(GLint x, GLint y)
{
// draw dot at integer point (x, y)
glBegin(GL_POINTS);
glVertex2i(x, y);
glEnd();
}
Figure 2.8. Encapsulating OpenGL details in a generic function drawDot()6.
The OpenGL State.
OpenGL keeps track of many state variables, such as the current size of points, the current color of
drawing, the current background color, etc. The value of a state variable remains active until a new value is
given. The size of a point can be set with glPointSize(), which takes one floating point argument. If its
argument is 3.0 the point is usually drawn as a square three pixels on a side. For additional details on this and
other OpenGL functions consult appropriate OpenGL documentation (some of which is on-line; see Appendix
1). The drawing color can be specified using
glColor3f(red, green, blue);
where the values of red, green, and blue vary between 0.0 and 1.0. For example, some of the colors listed in
Figure 1.3.24??? could be set using:
glColor3f(1.0,
glColor3f(0.0,
glColor3f(1.0,
glColor3f(1.0,

0.0,
0.0,
1.0,
1.0,

0.0);
0.0);
1.0);
0.0);

//
//
//
//

set
set
set
set

drawing
drawing
drawing
drawing

color
color
color
color

to
to
to
to

red
black
white
yellow

6 Using this function instead of the specific OpenGL commands makes a program more readable. It is not unusual to
build up a personal collection of such utilities.
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 7

The background color is set with glClearColor(red, green, blue, alpha), where alpha
specifies a degree of transparency and is discussed later (use 0.0 for now.) To clear the entire window to the
background color, use glClear(GL_COLOR_BUFFER_BIT). The argument GL_COLOR_BUFFER_BIT is
another constant built into OpenGL.
Establishing the Coordinate System.
Our method for establishing our initial choice of coordinate system will seem obscure here, but will become
clearer in the next chapter when we discuss windows, viewports, and clipping. Here we just take the few
required commands on faith. The myInit() function in Figure 2.9 is a good place to set up the coordinate
system. As we shall see later, OpenGL routinely performs a large number of transformations. It uses matrices
to do this, and the commands in myInit() manipulate certain matrices to accomplish the desired goal. The
gluOrtho2D() routine sets the transformation we need for a screen window that is 640 pixels wide by 480
pixels high.
void myInit(void)
{
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(0, 640.0, 0, 480.0);
}
Figure 2.9. Establishing a simple coordinate system.
Putting it together: A Complete OpenGL program.
Figure 2.10 shows a complete program that draws the lowly three dots of Figure 2.5. It is easily extended to
draw more interesting objects as we shall see. The initialization in myInit() sets up the coordinate system,
the point size, the background color, and the drawing color. The drawing is encapsulated in the callback
function myDisplay(). As this program is non-interactive, no other callback functions are used. glFlush()
is called after the dots are drawn to insure that all data is completely processed and sent to the display. This is
important in some systems that operate over a network: data is buffered on the host machine and only sent to
the remote display when the buffer becomes full or a glFlush() is executed.
#include <windows.h>
// use as needed for your system
#include <gl/Gl.h>
#include <gl/glut.h>
//<<<<<<<<<<<<<<<<<<<<<<< myInit >>>>>>>>>>>>>>>>>>>>
void myInit(void)
{
glClearColor(1.0,1.0,1.0,0.0);
// set white background color
glColor3f(0.0f, 0.0f, 0.0f);
// set the drawing color
glPointSize(4.0);
// a dot is 4 by 4 pixels
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(0.0, 640.0, 0.0, 480.0);
}
//<<<<<<<<<<<<<<<<<<<<<<<< myDisplay >>>>>>>>>>>>>>>>>
void myDisplay(void)
{
glClear(GL_COLOR_BUFFER_BIT);
// clear the screen
glBegin(GL_POINTS);
glVertex2i(100, 50);
// draw three points
glVertex2i(100, 130);
glVertex2i(150, 130);
glEnd();
glFlush();
// send all output to display
}
//<<<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char** argv)
{
glutInit(&argc, argv);
// initialize the toolkit
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB); // set display mode
glutInitWindowSize(640,480);
// set window size
glutInitWindowPosition(100, 150); // set window position on screen
glutCreateWindow("my first attempt"); // open the screen window
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 8

glutDisplayFunc(myDisplay);
myInit();
glutMainLoop();

// register redraw function

// go into a perpetual loop

}
Figure 2.10. A complete OpenGL program to draw three dots.
2.2.1. Drawing Dot Constellations.
A dot constellation is some pattern of dots or points. We describe several examples of interesting dot
constellations that are easily produced using the basic program in Figure 2.10. In each case the appropriate
function is named in glutDisplayFunc() as the callback function for the redraw event. You are strongly
encouraged to implement and test each example, in order to build up experience.
Example 2.2.1. The Big Dipper.
Figure 2.11 shows a pattern of eight dots representing the Big Dipper, a familiar sight in the night sky.

Figure 2.11. Two simple Dot Constellations.

The names and positions of the eight stars in the Big Dipper (for one particular view of the night sky), are
given by: {Dubhe, 289, 190}, {Merak, 320, 128}, {Phecda, 239, 67}, {Megrez, 194, 101}, {Alioth, 129, 83},
{Mizar, 75, 73}, {Alcor, 74, 74}, {Alkaid, 20, 10}. Since so few data points are involved it is easy to list
them explicitly, or hard-wire them into the code. (When many dots are to be drawn, it is more convenient
to store them in a file, and then have the program read them from the file and draw them. We do this in a later
chapter. ) These points can replace the three points specified in Figure 2.10. It is useful to experiment with
this constellation, trying different point sizes, as well as different background and drawing colors.
Example 2.2.2. Drawing the Sierpinski Gasket.
Figure 2.12 shows the Sierpinski gasket. Its dot constellation is generated procedurally, which means that
each successive dot is determined by a procedural rule. Although the rule here is very simple, the final pattern
is a fractal (see Chapter 8)! We first approach the rules for generating the Sierpinski gasket in an intuitive
fashion. In Case Study 2.2 we see that it is one example of an iterated function system.

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 9

Figure 2.12. The Sierpinski Gasket. (file: fig2.12.bmp)

The Sierpinski gasket is produced by calling drawDot() many times with dot positions (x0, y0), (x1, y1), (x2,
y2),.... determined by a simple algorithm. Denote the k-th point pk= (xk,yk). Each point is based on the
previous point pk-1.The procedure is:
1. Choose three fixed points T0, T1, and T2 to form some triangle, as shown in Figure 2.13a.
2. Choose the initial point p0 to be drawn by selecting one of the points T0, T1, and T2 at random.
Now iterate steps 3-5 until the pattern is satisfyingly filled in:
3. Chose one of the three points T0, T1, and T2 at random; call it T.
4. Construct the next point pk as the midpoint7 between T and the previously found point p . Hence
k-1

b).

a).

p
p

Figure 2.13. Building the Sierpinski gasket.

pk = midpoint of pk-1 and T,
5. Draw pk using drawDot().
Figure 2.13b shows a few iterations of this: Suppose the initial point p0 happens to be T0, and that T1 is chosen
next. Then p1 is formed so that it lies halfway between T0 and T1. Suppose T2 is chosen next, so p2 lies halfway
between p1 and T2. Next suppose T1 is chosen again, so p3 is formed as shown, etc. This process goes on
generating and drawing points (conceptually forever), and the pattern of the Sierpinski gasket quickly
emerges.
7 To find the midpoint between 2 points, say (3, 12) and (5, 37) simply average their x and y components individually:
add them and divide by 2. So the midpoint of (3,12) and (5,37) is ((3 + 5)/ 2, (12 + 37)/ 2) = (4, 24).
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 10

It is convenient to define a simple class GLintPoint that describes a point whose coordinates are integers8:
class GLintPoint{
public:
GLint x, y;
};
We then build and initialize an array of three such points T[0], T[1], and T[2] to hold the three corners
of the triangle using GLintPoint T[3]= {{10,10},{300,30},{200, 300}}. There is no
need to store each point pk in the sequence as it is generated, since we simply want to draw it and then move
on. So we set up a variable point to hold this changing point. At each iteration point is updated to hold
the new value.
We use i = random(3) to choose one of the points T[i] at random. random(3) returns one of the values
0, 1, or 2 with equal likelihood. It is defined as9
int random(int m) { return rand() % m; }
Figure 2.14 shows the remaining details of the algorithm, which generates 1000 points of the Sierpinski
gasket.

void Sierpinski(void)
{
GLintPoint T[3]= {{10,10},{300,30},{200, 300}};
int index = random(3);
// 0, 1, or 2 equally likely
GLintPoint point = T[index];
// initial point
drawDot(point.x, point.y);
// draw initial point
for(int i = 0; i < 1000; i++) // draw 1000 dots
{
index = random(3);
point.x = (point.x + T[index].x) / 2;
point.y = (point.y + T[index].y) / 2;
drawDot(point.x,point.y);
}
glFlush();
}
Figure 2.14. Generating the Sierpinski Gasket.
Example 2.2.3. Simple Dot Plots.
Suppose you wish to learn the behavior of some mathematical function f(x) as x varies. For example, how
does
-x

f(x) = e cos(2x)
vary for values of x between 0 and 4? A quick plot of f(x) versus x, such as that shown in Figure 2.15, can
reveal a lot.

8 If C rather than C++ is being used, a simple struct is useful here: typedef struct{GLint x,

y;}GLintPoint;
9 Recall that the standard function rand() returns a pseudorandom value in the range 0 to 32767. The modulo
function reduces it to a value in the range 0 to 2.
Computer Graphics Chap 2
09/14/99
1:02 PM
page 11

Figure 2.15. A dot plot of e-xcos(2 x) versus x. (file: fig2.15.bmp)

To plot this function, simply sample it at a collection of equispaced x-values, and plot a dot at each
coordinate pair (xi, f(xi)). Choosing some suitable increment, say 0.005, between consecutive x-values the
process is basically:
glBegin(GL_POINTS);
for(GLdouble x = 0; x < 4.0 ; x += 0.005)
glVertex2d(x, f(x));
glEnd();
glFlush();
But there is a problem here: the picture produced will be impossibly tiny because values of x between 0 and 4
map to the first four pixels at the bottom left of the screen window. Further, the negative values of f(.) will lie
below the window and will not be seen at all. We therefore need to scale and position the values to be plotted
so they cover the screen window area appropriately. Here we do it by brute force, in essence picking some
values to make the picture show up adequately on the screen. Later we develop a general procedure that copes
with these adjustments, the so-called procedure of mapping from world coordinates to window coordinates.
Scaling x: Suppose we want the range from 0 to 4 to be scaled so that it covers the entire width of the screen
window, given in pixels by screenWidth. Then we need only scale all x-values by screenWidth/4 ,
using
sx = x * screenWidth /4.0;
which yields 0 when x is 0 , and screenWidth when x is 4.0, as desired.
Scaling, and shifting y: The values of f(x) lie between 1.0 and 1.0, so we must scale and shift them as
well. Suppose we set the screen window to have height screenHeight pixels. Then to place the plot in the
center of the window scale by screenHeight / 2 and shift up by screenHeight / 2:
sy = (y + 1.0) * screenHeight / 2.0;
As desired, this yields 0 when y is 1.0, and screenHeight when y is 1.0.
Note that the conversions from x to sx, and from y to sy, are of the form:
sx = A * x + B
sy = C * y + D

Computer Graphics

(2.1)

Chap 2

09/14/99

1:02 PM

page 12

for properly chosen values of the constants A, B, C, and D. A and C perform scaling; B and D perform
shifting. This scaling and shifting is basically a form of affine transformation. We study affine
transformations in depth in Chapter 5. They provide a more consistent approach that maps any specified range
in x and y to the screen window.
We need only set the values of A, B, C, and D appropriately, and draw the dot -plot using:
GLdouble A, B, C, D, x;
A = screenWidth / 4.0;
B = 0.0;
C = screenHeight / 2.0;
D = C;
glBegin(GL_POINTS);
for(x = 0; x < 4.0 ; x += 0.005)
glVertex2d(A * x + B, C * f(x) + D);
glEnd();
glFlush();
Figure 2.16 shows the entire program to draw the dot plot, to illustrate how the various ingredients fit together. The
initializations are very similar to those for the program that draws three dots in Figure 2.10. Notice that the width and
height of the screen window are defined as constants, and used where needed in the code.
#include <windows.h> // use proper includes for your system
#include <math.h>
#include <gl/Gl.h>
#include <gl/glut.h>
const int screenWidth = 640;
// width of screen window in pixels
const int screenHeight = 480;
// height of screen window in pixels
GLdouble A, B, C, D; // values used for scaling and shifting
//<<<<<<<<<<<<<<<<<<<<<<< myInit >>>>>>>>>>>>>>>>>>>>
void myInit(void)
{
glClearColor(1.0,1.0,1.0,0.0);
// background color is white
glColor3f(0.0f, 0.0f, 0.0f);
// drawing color is black
glPointSize(2.0);
// a 'dot' is 2 by 2 pixels
glMatrixMode(GL_PROJECTION);
// set "camera shape"
glLoadIdentity();
gluOrtho2D(0.0, (GLdouble)screenWidth, 0.0, (GLdouble)screenHeight);
A = screenWidth / 4.0; // set values used for scaling and shifting
B = 0.0;
C = D = screenHeight / 2.0;
}
//<<<<<<<<<<<<<<<<<<<<<<<< myDisplay >>>>>>>>>>>>>>>>>
void myDisplay(void)
{
glClear(GL_COLOR_BUFFER_BIT);
// clear the screen
glBegin(GL_POINTS);
for(GLdouble x = 0; x < 4.0 ; x += 0.005)
{
Gldouble func = exp(-x) * cos(2 * 3.14159265 * x);
glVertex2d(A * x + B, C * func + D);
}
glEnd();
glFlush();
// send all output to display
}
//<<<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char** argv)
{
glutInit(&argc, argv);
// initialize the toolkit
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB); // set display mode
glutInitWindowSize(screenWidth, screenHeight); // set window size
glutInitWindowPosition(100, 150); // set window position on screen
glutCreateWindow("Dot Plot of a Function"); // open the screen window
glutDisplayFunc(myDisplay);
// register redraw function
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 13

myInit();
glutMainLoop();

// go into a perpetual loop

}
Figure 2.16. A complete program to draw the dot plot of a function.
Practice Exercise 2.2.2. Dot plots for any function f(). Consider drawing a dot plot of the function f(.) as in
Example 2.2.4, where it is known that as x varies from xlow to xhigh, f(x) takes on values between ylow to yhigh.
Find the appropriate scaling and translation factors so that the dots will lie properly in a screen window with
width W pixels and height H pixels.

2.3. Making Line-Drawings.

Hamlet: Do you see yonder cloud that's almost in shape of a camel?
Polonius: By the mass, and 'tis like a camel, indeed.
Hamlet: Methinks it is like a weasel.
William Shakespeare, Hamlet
As discussed in Chapter 1, line drawings are fundamental in computer graphics, and almost every graphics
system comes with driver routines to draw straight lines. OpenGL makes it easy to draw a line: use
GL_LINES as the argument to glBegin(), and pass it the two end points as vertices. Thus to draw a line
between (40,100) and (202,96) use:
glBegin(GL_LINES);
// use constant GL_LINES here
glVertex2i(40, 100);
glVertex2i(202, 96);
glEnd();
This code might be encapsulated for convenience in the routine drawLineInt():
void drawLineInt(GLint x1, GLint y1, GLint x2, GLint y2)
{
glBegin(GL_LINES);
glVertex2i(x1, y1);
glVertex2i(x2, y2);
glEnd();
}
and an alternate routine, drawLineFloat() could be implemented similarly (how?).
If more than two vertices are specified between glBegin(GL_LINES) and glEnd() they are taken in pairs
and a separate line is drawn between each pair. The tic-tac-toe board shown in Figure 2.17a would be drawn
using:
a). thin lines
b). thick lines
c). stippled lines

Figure 2.17. Simple picture built from four lines.

glBegin(GL_LINES);
glVertex2i(10, 20); // first horizontal line
glVertex2i(40, 20)
glVertex2i(20, 10); // first vertical line
glVertex2i(20, 40);
<four more calls to glVertex2i() here for other two lines>
glEnd();
glFlush();
OpenGL provides tools for setting the attributes of lines. A lines color is set in the same way as for points,
using glColor3f(). Figure 2.17b shows the use of thicker lines, as set by glLineWidth(4.0). The
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 14

default thickness is 1.0. Figure 2.17c shows stippled (dotted and dashed) lines. The details of stippling are
addressed in Case Study 2.5 at the end of this chapter.
2.3.1. Drawing Polylines and Polygons.
Recall from Chapter 1 that a polyline is a collection of line segments joined end to end. It is described by an
ordered list of points, as in:
p0 = (x0, y0), p1 = (x1, y1), ... , pn = (xn, yn).

(2.3.1)

In OpenGL a polyline is called a line strip, and is drawn by specifying the vertices in turn between
glBegin(GL_LINE_STRIP) and glEnd(). For example, the code:
glBegin(GL_LINE_STRIP);
glVertex2i(20,10);
glVertex2i(50,10);
glVertex2i(20,80);
glVertex2i(50,80);
glEnd();
glFlush();

// draw an open polyline

1:02 PM

page 18

2.3.3. Line Drawing using moveto() and lineto().

As we noted earlier a number of graphics systems provide line drawing tools based on the functions
moveto() and lineto(). These functions are so common it is important to be familiar with their use. We
shall fashion our own moveto() and lineto() that operate by calling OpenGL tools. In Chapter 3 we shall
also dive under the hood to see how you would build moveto() and lineto() based on first principles, if
a powerful library like OpenGL were not available.
Recall that moveto() and lineto() manipulate the position of a hypothetical pen, whose position is called
the current position, or CP. We can summarize the effects of the two functions as:
moveto(x, y):
lineto(x, y):

set CP to (x, y)
draw a line from C2970P to (x, y), and then update CP to (x, y)

A line from (x1, y1) to (x2, y2) is therefore drawn using the two calls moveto(x1, y1);lineto(x2,
y2). A polyline based on the list of points (x0, y0), (x1, y1), ... , (xn-1, yn-1) is easily drawn using:
moveto(x[0], y[0]);
for(int i = 1; i < n; i++)
lineto(x[i], y[i]);
It is straightforward to build moveto() and lineto() on top of OpenGL. To do this we must define and
maintain our own CP. For the case of integer coordinates the implementation shown in Figure 2.29 would do
the trick.
GLintPoint CP;
// global current position
//<<<<<<<<<<<<< moveto >>>>>>>>>>>>>>
void moveto(GLint x, GLint y)
{
CP.x = x; CP.y = y; // update the CP
}
//<<<<<<<<<<<< lineTo >>>>>>>>>>>>>>>>>
void lineto(GLint x, GLint y)
{
glBegin(GL_LINES); // draw the line
glVertex2i(CP.x, CP.y);
glVertex2i(x, y);
glEnd();
glFlush();
CP.x = x; CP.y = y; // update the CP
}
Figure 2.29. Defining moveto() and lineto() in OpenGL.
2.3.4. Drawing Aligned Rectangles.
A special case of a polygon is the aligned rectangle, so called because its sides are aligned with the
coordinate axes. We could create our own function to draw an aligned rectangle (how?), but OpenGL
provides the ready-made function:
glRecti(GLint x1, GLint y1, GLint x2, GLint y2);
// draw a rectangle with opposite corners (x1, y1) and (x2, y2);
// fill it with the current color;
This command draws the aligned rectangle based on two given points. In addition the rectangle is filled with
the current color. Figure 2.30 shows what is drawn by the code:

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 19

Figure 2.30. Two aligned rectangles filled with colors.

glClearColor(1.0,1.0,1.0,0.0); // white background
glClear(GL_COLOR_BUFFER_BIT); // clear the window
glColor3f(0.6,0.6,0.6);
// bright gray
glRecti(20,20,100,70);
glColor3f(0.2,0.2,0.2);
// dark gray
glRecti(70, 50, 150, 130);
glFlush();
Notice that the second rectangle is painted over the first one. We examine other drawing modes in Chapter 10.
Figure 2.31 shows two further examples. Part a) is a flurry of randomly chosen aligned rectangles, that might be
generated by code such as10:
a).

b).

Figure 2.31. a). Random Flurry of rectangles. b). a checkerboard.

void drawFlurry(int num, int numColors, int Width, int Height)
// draw num random rectangles in a Width by Height rectangle
{
for (int i = 0; i < num; i++)
{
GLint x1 = random(Width);
// place corner randomly
GLint y1 = random(Height);
GLint x2 = random(Width);
// pick the size so it fits
GLint y2 = random(Height);
GLfloat lev = random(10)/10.0;
// random value, in range 0 to 1
glColor3f(lev,lev,lev);
// set the gray level
glRecti(x1, y1, x2, y2);
// draw the rectangle
}
glFlush();
}
Part b) is the familiar checkerboard, with alternating gray levels. The exercises ask you to generate it.
2.3.5. Aspect Ratio of an Aligned Rectangle.

10 Recall that random(N) returns a randomly-chosen value between 0 and N - 1 (see Appendix 3).

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 20

The principal properties of an aligned rectangle are its size, position, color, and shape. Its shape is
embodied in its aspect ratio, and we shall be referring to the aspect ratios of rectangles throughout the book.
The aspect ratio of a rectangle is simply the ratio of its width to its height11:

aspect ratio =

width
height

(2.2)

Rectangles with various aspect ratios are shown in Figure 2.32..

Figure 2.32. Examples of aspect ratios of aligned rectangles.

Rectangle A has the shape of a piece of 8.5 by 11 inch paper laid on its side in the so-called landscape
orientation (i.e. width larger than height). It has an aspect ratio of 1.294. Rectangle B has the aspect ratio of a
television screen, 4/3, and C is the famous golden rectangle described in Case Study 2.3. Its aspect ratio is
close to = 1.618034. Rectangle D is a square with aspect ratio equal to 1, and E has the shape of a piece of
standard paper in portrait orientation, with an aspect ratio of .7727. Finally, F is tall and skinny with an
aspect ratio of 1/.
Practice Exercises.
2.3.1. Drawing the checkerboard. (Try your hand at this before looking at the answers.)
Write the routine checkerboard(int size) that draws the checkerboard shown in Figure 2.31b. Place
the checkerboard with its lower left corner at (0,0). Each of the 64 squares has length size pixels. Choose
two nice colors for the squares. Solution: The ij-th square has lower left corner at (i*size, j*size) for i =
0,..,7 and j = 0..7. The color can be made to alternate between (r1, g1, b1) and (r2, g2, b2) using
if((i + j)%2 ==0) // if i + j is even
glColor3f( r1, g1, b1);
else
glColor3f(r2, g2, b2);
2.3.2. Alternative ways to specify a rectangle. An aligned rectangle can be described in other ways than by
two opposite corners. Two possibilities are:
* its center point, height, and width;
* its upper left corner, width, and aspect ratio.
Write functions drawRectangleCenter() and drawRectangleCornerSize() that pass these
alternative parameters.
2.3.3. Different Aspect Ratios. Write a short program that draws a filled rectangle of aspect ratio R, where R
is specified by the user. Initialize the display to a drawing space of 400 by 400. Arrange the size of the
rectangle so that it is as large as possible. That is, if R > 1 it spans across the drawing space, and if R < 1 it
spans from top to bottom.
2.3.4. Drawing the Parametrized house. Fill in the details of parametrizedHouse() in Figure 2.25 so
that the door, window, and chimney are drawn in their proper proportions for given values of height and
width.

11Alert! Some authors define it as height / width.

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 21

2.3.5. Scaling and positioning a figure using parameters. Write the function void
drawDiamond(GLintPoint center, int size) that draws the simple diamond shown in Figure
2.33, centered at center, and having size size.
x

Figure 2.33. A simple diamond.

Use this function to draw a flurry of diamonds as suggested in Figure 2.34.

Figure 2.34. A flurry of diamonds.

2.3.6. Filling Polygons.
So far we can draw unfilled polygons in OpenGL, as well as aligned rectangles filled with a single solid color.
OpenGL also supports filling more general polygons with a pattern or color. The restriction is that the
polygons must be convex.
Convex polygon:
a polygon is convex if a line connecting any two points of the polygon lies entirely within the polygon.
Several polygons are shown in Figure 2.35. Of these only D, E, and F are convex. (Check that the definition
of convexity is upheld for each of these polygons.) D is certainly convex: all triangles are. A is not even
simple (recall Chapter 1) so it cannot be convex. Both B and C bend inward at some point. (Find two points
on B such that the line joining them does not lie entirely inside B.)

Figure 2.35. Convex and non-convex polygons.

To draw a convex polygon based on vertices (x0, y0), (x1, y1), , (xn, yn) use the usual list of vertices, but place
them between a glBegin(GL_POLYGON) and an glEnd():
glBegin(GL_POLYGON);
glVertex2f(x0, y0);
glVertex2f(x1, y1);
. . .
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 22

glVertex2f(xn, yn);
glEnd();
It will be filled in the current color. It can also be filled with a stipple pattern see Case Study 2.5, and later
we will paint images into polygons as part of applying a texture.
Figure 2.36 shows a number of filled convex polygons. In Chapter 10 we will examine an algorithm for filling
any polygon, convex or not.

Figure 2.36. Several filled convex polygons.

1:02 PM

page 25

Figure 2.39. Interactive creation of a polyline.

Figure 2.40 shows one possible implementation. Note that last keeps track of the last index used so far in
the array List[]; it is incremented as each new point is clicked, and set to 1 to make the lists empty. If it
were desirable to make use of the points in List outside of myMouse(), the variable List could be made
global.
void myMouse(int button, int state, int x, int y)
{
#define NUM 20
static GLintPoint List[NUM];
static int last = -1;
// last index used so far
// test for mouse button as well as for a full array
if(button == GLUT_LEFT_BUTTON && state == GLUT_DOWN && last < NUM -1)
{
List[++last].x = x;
// add new point to list
List[ last].y = screenHeight - y; // window height is 480
glClear(GL_COLOR_BUFFER_BIT);
// clear the screen
glBegin(GL_LINE_STRIP);
// redraw the polyline
for(int i = 0; i <= last; i++)
glVertex2i(List[i].x, List[i].y);
glEnd();
glFlush();
}
else if(button == GLUT_RIGHT_BUTTON && state == GLUT_DOWN)
last = -1;
// reset the list to empty
}
Figure 2.40. A polyline drawer based on mouse clicks.
Mouse motion.
An event of a different type is generated when the mouse is moved (more than some minimal distance) while
some button is held down. The callback function, say myMovedMouse(), is registered with this event using
glutMotionFunc(myMovedMouse);
The callback function must take two parameters and have the prototype: myMovedMouse(int x, int y);
The values of x and y are of course the position of the mouse when the event occurred.
Example 2.4.4. Freehand drawing with a fat brush.
Suppose we want to create a curve by sweeping the mouse along some trajectory with a button held down. In
addition we want it to seem that the drawing brush has a square shape. This can be accomplished by
designing myMovedMouse() to draw a square at the current mouse position:
void myMovedMouse(int mouseX, int mouseY)
{
GLint x = mouseX;
//grab the mouse position
GLint y = screenHeight mouseY;
// flip it as usual
GLint brushSize = 20;
glRecti(x,y, x + brushSize, y + brushSize);
glFlush();
}
2.4.2. Keyboard interaction.
As mentioned earlier, pressing a key on the keyboard queues a keyboard event. The callback function
myKeyboard() is registered with this type of event through glutKeyboardFunc(myKeyboard).
It must have prototype:
void myKeyboard(unsigned int key, int x, int y);

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 26

The value of key is the ASCII value12 of the key pressed. The values x and y report the position of the mouse
at the time that the event occurred. (As before y measures the number of pixels down from the top of the
window.)
The programmer can capitalize on the many keys on the keyboard to offer the user a large number of choices
to invoke at any point in a program. Most implementations of myKeyboard() consist of a large switch
statement, with a case for each key of interest. Figure 2.41 shows one possibility. Pressing p draws a dot at
the mouse position; pressing the left arrow key adds a point to some (global) list, but does no drawing13;
pressing E exits from the program. Note that if the user holds down the p key and moves the mouse around
a rapid sequence of points is generated to make a freehand drawing.
void myKeyboard(unsigned char theKey, int mouseX, int mouseY)
{
GLint x = mouseX;
GLint y = screenHeight - mouseY; // flip the y value as always
switch(theKey)
{
case p:
drawDot(x, y);
// draw a dot at the mouse position
break;
case GLUT_KEY_LEFT: List[++last].x = x; // add a point
List[ last].y = y;
break;
case E:
exit(-1);
//terminate the program
default:
break;
// do nothing
}
}
Figure 2.41. An example of the keyboard callback function.

2.5. Summary
The hard part in writing graphics applications is getting started: pulling together the hardware and software
ingredients in a program to make the first few pictures. The OpenGL application programmer interface (API)
helps enormously here, as it provides a powerful yet simple set of routines to make drawings. One of its great
virtues is device independence, which makes it possible to write programs for one graphics environment, and
use the same program without changes in another environment.
Most graphics applications are written today for a windows-based environment. The program opens a window
on the screen that can be moved and resized by the user, and it responds to mouse clicks and key strokes. We
saw how to use OpenGL functions that make it easy to create such a program.
Primitive drawing routines were applied to making pictures composed of dots, lines, polylines, and polygons,
and were combined into more powerful routines that form the basis of one's personal graphics toolkit. Several
examples illustrated the use of these tools, and described methods for interacting with a program using the
keyboard and mouse. The Case studies presented next offer additional programming examples that explore
deeper into the topics discussed so far, or branch out to interesting related topics.

2.6. Case Studies.

It is best while using this text to try out new ideas as they are introduced, to solidify the ideas presented. This
is particularly true in the first few chapters, since getting started with the first graphics programs often
presents a hurdle. To focus this effort, each chapter ends with some Case Studies that describe programming
projects that are both interesting in themselves, and concentrate on the ideas developed in the chapter.

12 ASCII stands for American Standard Code for Information Interchange. Tables of ASCII values are readily

available on the internet. Also see ascii.html in the web site for this book.
13 Names for the various special keyboard keys, such as the function keys, arrow keys, and home, may be found
in the include file glut.h.
Computer Graphics Chap 2
09/14/99
1:02 PM
page 27

Some of the Case Studies are simple exercises that only require fleshing out some pseudocode given in the
text, and then running the program through its paces. Others are much more challenging, and could be the
basis of a major programming project within a course. It is always difficult to judge how much time someone
else will need to accomplish any project. The Level of Effort that accompanies each Case Study is a rough
guess at best.
Level of Effort:
I: a simple exercise. It could be assigned for the next class.
II: an intermediate exercise. It probably needs several days for completion14.
III: An advanced exercise. It would probably be assigned for two weeks or so ahead.

2.6.1. Case Study 2.1. Pseudo random Clouds of Dots.

(Level of Effort: II) The random number generator (RNG) random(N) (see Appendix 3) produces a value
between 0 and N-1 each time it is called. It uses the standard C++ function rand() to generate values. Each
value appears to be randomly selected, and to have no relation to its predecessors.
In fact the successive numbers that rand() produces are not generated randomly at all, but rather through a
very regular mechanism where each number ni is determined from its predecessor ni-1 by a specific formula.
A typical formula is:
ni = (ni-1 * A + B) mod N

(2.3)

where A, B, and N are suitably chosen constants. One set of numbers that works fairly well is: A =
1103515245, B = 12345, and N = 32767. Multiplying ni-1 by A and adding B forms a large value, and the
modulo operation brings the value into the range 0 to N-1. The process begins with some seed value chosen
for n0.
Because the numbers only give an appearance of randomness they are called pseudo random numbers. The
choices of the values for A, B, and N are very important, and slightly different values give rise to very
different characteristics in the sequence of numbers. More details can be found in [knuth, weiss98] .
Scatter Plots.
Some experiments yield data consisting of many pairs of numbers (ai, bi), and the goal is to infer visually how
the a-values and b-values are related. For instance, a large number of people are measured, and one
wonders if there is a strong correlation between a persons height and weight.
A scatter plot can be used to give visual insight into the data. The data for each person is plotted as a dot at
position (height, weight) so only the drawDot() tool is needed. Figure 2.42 shows an example. It suggests
that a persons height and weight are roughly linearly related, although some people (such as A) are
idiosyncratic, being very tall yet quite light.

Figure 2.42. A scatter plot of peoples height versus weight.

A day of programming means several two hour sessions, with plenty of thinking (and resting) time
between sessions. It also assumes a reasonably skilled programmer (with at least two semesters of
programming in hand), who is familiar with the idiosyncrasies of the language and the platform being used. It
does not allow for those dreadful hours we all know too well of being stuck with some obscure bug that
presents a brick wall of frustration until it is ferreted out and squashed.
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 28

Here we use scatter plots to visually test the quality of a random number generator. Each time the function
random(N)is called it returns a value in the range 0..N - 1 that is apparently chosen at random, unrelated to
values previously returned from random(N). But are successive values truly unrelated?
One simple test builds a scatter plot based on pairs of successive values returned by random(N). It calls
random(N) twice in succession, and plots the first value against the second. This can be done using
drawDot():
for(int i = 0; i < num; i++)
drawDot(random(N), random(N));
or in "raw" OpenGL by placing the for loop between glBegin() and glEnd():
glBegin(GL_POINTS);
for(int i = 0; i < num; i++)
glVertex2i(random(N), random(N));
glEnd();

// do it num times

It is more efficient to do it the second way, which avoids the overhead associated with making many calls to
glBegin() and glEnd().
Figure 2.43 shows a typical plot that might result. There should be a uniform density of dots throughout the
square, to reassure you that the values 0..N-1 occur with about equal likelihood, and that there is no
discernible dependence between one value and its successor.

Figure 2.43. A constellation of 500 random dots.

Figure 2.44 shows what can happen with an inferior RNG. In a) there is too high a density of certain values,
so the distribution is not uniform in 0..N-1. In b) there is high correlation between a number and its successor:
when one number is large the other tends to be small. And c) shows perhaps the worst situation of all: after a
few dozen values have been generated the pattern repeats, and no new dots are generated!

a).

c).

b).

Figure 2.44. Scatter plots for inferior random number generators.

Plotting dot constellations such as these can provide a rough first check on the uniformity of the numbers
generated. It is far from a thorough test, however [knuth].
Write a program that produces random dot plots, using some different RNGs to produce the (x, y) pairs. Try
different constants A, B, and N in the basic RNG and see the effect this has on the dot constellations.
(Warning: if the dot constellation suddenly freezes such that no new dots appear, it may be that the pattern
of numbers is simply repeating.)

2.6.2. Case Study 2.2. Introduction to Iterated Function Systems.

2.6.3. Case Study 2.3. The Golden Ratio and Other Jewels.
(Level of Effort: I.) The aspect ratio of a rectangle is an important attribute. Over the centuries, one aspect
ratio has been particularly celebrated for its pleasing qualities in works of art: that of the golden rectangle.
The golden rectangle is considered as the most pleasing of all rectangles, being neither too narrow nor too
squat. It figures in the Greek Parthenon (see Figure 2.48), Leonardo da Vinci's Mona Lisa, Salvador Dali's
The Sacrament of the Last Supper, and in much of M. C. Escher's works.
old Fig 3.18 (picture of Greek Parthenon inside golden rectangle)
Figure 2.48. The Greek Parthenon fitting within a Golden Rectangle.
The golden rectangle is based on a fascinating quantity, the golden ratio = 1.618033989.. The value
appears in a surprising number of places in computer graphics.
Figure 2.49 shows a golden rectangle, with sides of length and 1. Its shape has the unique property that if a
square is removed from the rectangle, the piece that remains will again be a golden rectangle! What value
must have to make this work? Note in the figure that the smaller rectangle has height 1 and so to be golden
must have width 1/. Thus

Figure 2.49. The Golden Rectangle.

=1+ 1

(2.8)

which is easily solved to yield

= 1+ 5 = 1.618033989...
2

(2.9)

This is approximately the aspect ratio of a standard 3-by-5 index card. From Equation 2.8 we see also that if 1
is subtracted from the reciprocal of is obtained: 1/ = .618033989. . . . This is the aspect ratio of a golden
rectangle lying on its short end.
The number is remarkable mathematically in many ways, two favorites being

= 1 + 1 + 1 + 1 + ...

(2.10)

and

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 33

=1+

1
1

1+
1+

1
1 + ...

(2.11)

These both are easy to prove (how?) and display a pleasing simplicity in the use of the single digit 1.
The idea that the golden rectangle contains a smaller version of itself suggests a form of infinite regression
of figureswithin figureswithin figures ad infinitum. Figure 2.50 demonstrates this. Keep removing
squares from each remaining golden rectangle.

Figure 2.50. Infinite regressions of the golden rectangle.

Write an application that draws the regression of golden rectangles centered in a screen window 600 pixels
wide by 400 pixels high. (First determine where and how big the largest golden rectangle is that will fit in this
screen window. Your picture should regress down until the smallest rectangle is about one pixel in size.
There is much more to be said about the golden ratio, and many delights can be found in [Gardner 1961],
[Hill 1978], [Huntley 1970], and [Ogilvy 1969]. For instance, in the next chapter we see golden pentagrams,
and in Chapter 9 we see that two of the platonic solids, the dodecahedron and the icosahedron, contain three
mutually perpendicular golden rectangles!
Practice exercises.
2.6.2. Other golden things. Equation 2.10 shows as a repeated square root involving the number 1. What
is the value of:

W = k+ k+ k + k +

2.6.3. On and Golden Rectangles. a). Show the validity of Equations 2.10 and 2.11.
b). Find the point at which the two dotted diagonals shown in Figure 2.50 lie, and show that this is the point to
which the sequence of golden rectangles converges.
c). Use Equation 2.8 to derive the relationship:

2 + 1 = 3
2

(2.12)
2.6.4. Golden orbits. The expressions in Equations 2.10 and 2.11 show that the golden ratio is the limiting
value of applying certain functions again and again. The first function is f(.) = 1 + (.) . What is the second
function? Viewing these expressions in terms of iterated functions systems, is seen to be the value to which
orbits converge for some starting values. (The starting value is hidden in the ... of the expressions.) Explore
with a hand calculator what starting values one can use and still have the process converge to .

2.6.4. Case Study 2.4. Building and Using Polyline Files.

(Level of Effort: II) Complex pictures such as Figure 2.21 are based on a large collection of polylines. The
data for the polylines are typically stored in a file, so the picture can be reconstructed at a later time by
reading the polylines into a program and redrawing each line. A reasonable format for such a file was
described in Section 2.3.1, and the routine drawPolyLineFile() was described that does the drawing.
The file, dino.dat, that stores the dinosaur in Figure 2.21 is available as dino.dat on the web site for
this book (see the preface) Other polyline files are also available there.
a). Write a program that reads polyline data from a file and draws each polyline in turn. Generate at least one
interesting polyline file of your own on a text processor, and use your program to draw it.
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 34

b). Extend the program in the previous part to accept some other file formats. For instance, have it accept
differentially coded x- and y- coordinates. Here the first point (x1, y1) of each polyline is encoded as above,
but each remaining one (xi, yi) is encoded after subtracting the previous point from it: the file contains (xi - xi, yi - yi-1). In many cases there are fewer significant digits in the difference than in the original point values,
1
allowing more compact files. Experiment with this format.
c). Adapt the file format above so that a color value is associated with each polyline in the file. This color
value appears in the file on the same line as the number of points in the associated polyline. Experiment with
several polyline files.

d). Adjust the polyline drawing routine so that it draws a closed polygon when a minus sign precedes the
number of points in a polyline, as in:
-3
0
35
57
5
0
12
23
.. etc

<-- negative: so this is a polygon

first point in this polygon
second point in this polygon
also connect this to the first point of the polygon
<-- positive, so leave it open as usual

0
3
8
1
21
34

The first polyline is drawn as a triangle: its last point is connected to its first..

2.6.5. Case Study 2.5. Stippling of Lines and Polygons.

(Level of Effort: II) You often want a line to be drawn with a dot-dash pattern, or to have a polygon filled
with a pattern representing some image. OpenGL provides convenient tools to do this.
Line Stippling.
It is straightforward to define a stipple pattern for use with line drawing. Once the pattern is specified, it is
applied to subsequent line drawing as soon as it is enabled with:
glEnable(GL_LINE_STIPPLE);
and until it is disabled with
glDisable(GL_LINE_STIPPLE).
The function
glLineStipple(GLint factor, GLushort pattern);
defines the stipple pattern. The value of pattern (which is of type GLushort an unsigned 16-bit
quantity) is a sequence of 0s and 1s that define which dots along the line are drawn: a 1 prescribes that a dot
is drawn; a 0 prescribes that it is not. The pattern is repeated as many times as necessary to draw the desired
line. Figure 2.51 shows several examples. The pattern is given compactly in hexadecimal notation. For
example, 0xEECC specifies the bit pattern 1110111011001100. The variable factor specifies how much to
enlarge pattern: each bit in the pattern is repeated factor times. For instance, the pattern 0xEECC
with factor 2 yields: 11111100111111001111000011110000. When drawing a stippled polyline, as with
glBegin(GL_LINE_STRIP); glVertex*(); glVertex*(); glVertex*(); . . .
glEnd();, the pattern continues from the end of one line segment to the beginning of the next, until the
glEnd().
pattern factor
resulting stipple
0x00FF
1
..
..
..
..
..
0x00FF
2
.
.
.
0xAAAA 1
.....................................
0xAAAA 2
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 35

0xEECC
1
.. .. .. .. .. .. .. .. .. ..
Figure 2.51. Example stipple patterns.
Write a program that allows the user to type in a pattern (in hexadecimal notation) and a value for factor,
and draws stippled lines laid down with the mouse.
Polygon Stippling.
It is also not difficult to define a stipple pattern for filling a polygon: but there are more details to cope with.
After the pattern is specified, it is applied to subsequent polygon filling once it is enabled with
glEnable(GL_POLYGON_STIPPLE), until disabled with glDisable(GL_POLYGON_STIPPLE).
The function
glPolygonStipple(const GLubyte * mask);
attaches the stipple pattern to subsequently drawn polygons, based on a 128 byte array mask[]. These 128
bytes provide the bits for a bitmask that is 32 bits wide and 32 bits high. The pattern is tiled throughout the
polygon (which is invoked with the usual glBegin(GL_POLYGON); glVertex*();. . .;
glEnd();).The pattern is specified by an array definition such as:
GLubyte mask[] = {0xff, 0xfe, 0x34, ... };
The first four bytes prescribe the 32 bits across the bottom row, from left to right; the next 4 bytes give the next
row up, etc. Figure 2.52 shows the result of filling a specific polygon with a fly pattern specified in the
OpenGL red book [woo97].

Figure 2.52. An example stippled polygon.

Write a program that defines an interesting stipple pattern for a polygon. It then allows the user to lay down a
sequence of convex polygons with the mouse, and each polygon is filled with this pattern.

2.6.6. Case Study 2.6. Polyline Editor.

(Level of Effort: III) Drawing programs often allow one to enter polylines using a mouse, and then to edit the
polylines until they present the desired picture. Figure 2.53a shows a house in the process of being drawn: the
user has just clicked at the position shown, and a line has been drawn from the previous point to the mouse
point.

a).Add points

b).Move a point

c).Delete a point

Figure 2.53. Creating and editing polylines.

Figure 2.53b shows the effect of moving a point. The user positions the mouse cursor near some polyline
vertex, presses down the mouse button, and drags the chosen point to some other location before releasing
the mouse button. Upon release, the previous lines connected to this point are erased, and new lines are drawn
to it.
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 36

Figure 2.53c shows how a point is deleted from a polyline. The user clicks near some polyline vertex, and the
two line segments connected to it are erased. Then the two other endpoints of the segments just erased are
connected with a line segment.
Write and exercise a program that allows the user to enter and edit pictures made up of as many as 60
polylines. The user interacts by pressing keyboard keys and pointing/clicking with the mouse. The
functionality of the program should include the actions:
begin
delete
move
refresh
quit

(b):
(d):
(m):
(r):
(q):

(create a new polyline)

(delete the next point pointed to)
(drag the point pointed to new location)
(erase the screen and redraw all the polylines)
(exit from the program)

A list of polylines can be maintained in an array such as: GLintPointArray polys[60]. The verb
begin, activated by pressing the key b, permits the user to create a new polyline, which is stored in the
first available slot in array polys. The verb delete requires that the program identify which point of
which polyline lies closest to the current mouse point. Once identified, the previous and next vertices in
the chosen polyline are found. The two line segments connected to the chosen vertex are erased, and the
previous and next vertices are joined with a line segment. The verb move finds the vertex closest to the
current mouse point, and waits for the user to click the mouse a second time, at which point it moves the
vertex to this new point.
What other functions might you want in a polyline editor? Discuss how you might save the array of polylines
in a file, and read it in later. Also discuss what a reasonable mechanism might be for inserting a new point
inside a polyline.

2.6.7. Case Study 2.7. Building and Running Mazes.

(Level of Effort: III) The task of finding a path through a maze seems forever fascinating (see [Ball and
Coxeter 1974]). You can generate an elaborate maze on a computer, and use graphics to watch it be traversed.
Figure 2.54 shows a rectangular maze having 100 rows and 150 columns. The goal is to find a path from the
opening at the left edge to the opening at the right edge. Although you can traverse it manually by trial and
error, it's more interesting to develop an algorithm to do it automatically.
use 1st Ed. Figure 3.38
Figure 2.54. A maze.
Write and exercise a program that a). generates and displays a rectangular maze of R rows and C columns,
and b). finds (and displays) the path from start to end. The mazes are generated randomly but must be proper;
that is, every one of the R-by-C cells is connected by a unique, albeit tortuous, path to every other cell. Think
of a maze as a graph, as suggested by Figure 2.55. A node of the graph corresponds to each cell where either a
path terminates or two paths meet, and each path is represented by a branch. So a node occurs at every cell for
which there is a choice of which way to go next. For instance, when Q is reached there are three choices,
whereas at M there are only two. The graph of a proper maze is acyclic and has a tree structure.
use 1st Ed. Figure 3.39
Figure 2.55. A simple maze and its graph.
How should a maze be represented? One way is to state for each cell whether its north wall is intact and its
east wall is intact, suggesting the following data structure:
char northWall[R][C], eastWall[R][C];
If northwall[i][j] is 1, the ij-th cell has a solid upper wall; otherwise the wall is missing. The 0-th row
is a phantom row of cells below the maze whose north walls comprise the bottom edge. Similarly,
eastwall[i][0] specifies where any gaps appear in the left edge of the maze.
Computer Graphics

Chap 2

09/14/99

1:02 PM

page 37

Generating a Maze. Start with all walls intact so that the maze is a simple grid of horizontal and vertical
lines. The program draws this grid. An invisible mouse whose job is to eat through walls to connect
adjacent cells, is initially placed in some arbitrarily chosen cell. The mouse checks the four neighbor cells
(above, below, left, and right) and for each asks whether the neighbor has all four walls intact. If not, the cell
has previously been visited and so is already on some path. The mouse may detect several candidate cells that
haven't been visited: It chooses one randomly and eats through the connecting wall, saving the locations of
the other candidates on a stack. The eaten wall is erased, and the mouse repeats the process. When it becomes
trapped in a dead endsurrounded by visited cellsit pops an unvisited cell and continues. When the stack is
empty, all cells in the maze have been visited. A start and end cell is then chosen randomly, most likely
along some edge of the maze. It is delightful to watch the maze being formed dynamically as the mouse eats
through walls. (Question: Might a queue be better than a stack to store candidates? How does this affect the
order in which later paths are created?)
Running the Maze. Use a backtracking algorithm. At each step, the mouse tries to move in a random
direction. If there is no wall, it places its position on a stack and moves to the next cell. The cell that the
mouse is in can be drawn with a red dot. When it runs into a dead end, it can change the color of the cell to
blue and backtrack by popping the stack. The mouse can even put a wall up to avoid ever trying the dead-end
cell again.
Addendum: Proper mazes aren't too challenging because you can always traverse them using the shoulderto-the-wall rule. Here you trace the maze by rubbing your shoulder along the left-hand wall. At a dead end,
sweep around and retrace the path, always maintaining contact with the wall. Because the maze is a tree,
you will ultimately reach your destination. In fact, there can even be cycles in the graph and you still always
find the end, as long as both the start and the end cells are on outer boundaries of the maze (why?). To make
things more interesting, place the start and end cells in the interior of the maze and also let the mouse eat
some extra walls (maybe randomly 1 in 20 times). In this way, some cycles may be formed that encircle the
end cell and defeat the shoulder method.

2.9. For Further reading

Several books provide an introduction to using OpenGL. The OpenGL Programming Guide by Woo,
Neider, and Davis [woo97] is an excellent source. There is also a wealth of information available on the
internet. See, for instance, the OpenGL repository at: http://www.opengl.org/, and the complete manual for
openGL at http://www.sgi.com/software/opengl/manual.html.

Computer Graphics

Chap 2

09/14/99

1:02 PM

page 38

(For ECE660 - Fall, 1999)

CHAPTER 3. More Drawing Tools.

Computers are useless.
They can only give you answers.
Pablo Picasso
Even if you are on the right track, youll
get run over if you just sit there.
Will Rogers

Goals of the Chapter

Introduce viewports and clipping

Develop the window to viewport transformation
Develop a classical clipping algorithm
Create tools to draw in world coordinates
Develop a C++ class to encapsulate the drawing routines
Develop ways to select windows and viewports for optimum viewing
Draw complex pictures using relative drawing, and turtle graphics
Build figures based on regular polygons and their offspring
Draw arcs and circles.
Describe parametrically defined curves and see how to draw them.

Preview.
Section 3.1 introduces world coordinates and the world window. Section 3.2 describes the window to viewport
transformation. This transformation simplifies graphics applications by letting the programmer work in a
reasonable coordinate system, yet have all pictures mapped as desired to the display surface. The section also
discusses how the programmer (and user) choose the window and viewport to achieve the desired drawings. A
key property is that the aspect ratios of the window and viewport must agree, or distortion results. Some of the
choices can be automated. Section 3.3 develops a classical clipping algorithm that removes any parts of the
picture that lie outside the world window.
Section 3.4 builds a useful C++ class called Canvas that encapsulates the many details of initialization and
variable handling required for a drawing program. Its implementation in an OpenGL environment is developed.
A programmer can use the tools in Canvas to make complex pictures, confident that the underlying data is
protected from inadvertent mishandling.
Section 3.5 develops routines for relative drawing and turtle graphics that add handy methods to the
programmers toolkit. Section 3.6 examines how to draw interesting figures based on regular polygons, and
Section 3.7 discusses the drawing of arcs and circles. The chapter ends with several Case Studies, including the
development of the Canvas class for a non-OpenGL environment, where all the details of clipping and the
window to viewport transformation must be explicitly developed.
Section 3.8 describes different representations for curves, and develops the very useful parametric form, that
permits straightforward drawing of complex curves. Curves that reside in both 2D space and 3D space are
considered.

3.1. Introduction.
It is as interesting and as difficult to say a thing well as to paint it.
Vincent Van Gogh
In Chapter 2 our drawings used the basic coordinate system of the screen window: coordinates that are
essentially in pixels, extending from 0 to some value screenWidth 1 in x, and from 0 to some value
screenHeight 1 in y. This means that we can use only positive values of x and y, and the values must
extend over a large range (several hundred pixels) if we hope to get a drawing of some reasonable size.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 1

In a given problem, however, we may not want to think in terms of pixels. It may be much more natural to think
in terms of x varying from, say, -1 to 1, and y varying from 100.0 to 20.0. (Recall how awkward it was to scale
and shift values when making the dot plots in Figure 2.16.) Clearly we want to make a separation between the
values we use in a program to describe the geometrical objects and the size and position of the pictures of them
on the display.
In this chapter we develop methods that let the programmer/user describe objects in whatever coordinate system
best fits the problem at hand, and to have the picture of the object automatically scaled and shifted so that it
comes out right in the screen window. The space in which objects are described is called world coordinates.
It is the usual Cartesian xy-coordinate system used in mathematics, based on whatever units are convenient.
We define a rectangular world window1 in these world coordinates. The world window specifies which part of
the world should be drawn. The understanding is that whatever lies inside the window should be drawn;
whatever lies outside should be clipped away and not drawn.
In addition, we define a rectangular viewport in the screen window on the screen. A mapping (consisting of
scalings and shiftings) between the world window and the viewport is established so that when all the objects in
the world are drawn, the parts that lie inside the world window are automatically mapped to the inside of the
viewport. So the programmer thinks in terms of looking through a window at the objects being drawn, and
placing a snapshot of whatever is seen in that window into the viewport on the display. This window/viewport
approach makes it much easier to do natural things like zooming in on a detail in the scene, or panning
around a scene.
We first develop the mapping part that provides the automatic change of coordinates. Then we see how clipping
is done.

3.2. World Windows and Viewports.

We use an example to motivate the use of world windows and viewports. Suppose you want to examine the
nature of a certain mathematical function, the sinc function famous in the signal processing field. It is defined
by

sinc( x ) =

sin(x )
x

(3.1)

You want to know how it bends and wiggles as x varies. Suppose you know that as x varies from - to the
value of sinc(x) varies over much of the range 1 to 1, and that it is particularly interesting for values of x near
0. So you want a plot that is centered at (0, 0), and that shows sinc(x) for closely spaced x-values between, say,
4.0 to 4.0. Figure 3.1 shows an example plot of the function. It was generated using the simple OpenGL
display function (after a suitable world window and viewport were specified, of course):

Figure 3.1. A plot of the sinc function.

void myDisplay(void)
1 As mentioned, the term window has a bewildering set of meanings in graphics, which often leads to confusion.
We will try to keep the different meanings clear by saying "world window", screen window", etc., when necessary.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 2

{
glBegin(GL_LINE_STRIP);
for(GLfloat x = -4.0; x < 4.0; x += 0.1)
{
GLfloat y = sin(3.14159 * x) / (3.14159 * x);
glVertex2f(x, y);
}
glEnd();
glFlush();
}
Note that the code in these examples operates in a natural coordinate system for the problem: x is made to vary
in small increments from 4.0 to 4.0. The key issue here is how the various (x, y) values become scaled and
shifted so that the picture appears properly in the screen window.
We accomplish the proper scaling and shifting by setting up a world window and a viewport, and establishing a
suitable mapping between them. The window and viewport are both aligned rectangles specified by the
programmer. The window resides in world coordinates. The viewport is a portion of the screen window. Figure
3.2 shows an example world window and viewport. The notion is that whatever lies in the world window is
scaled and shifted so that it appears in the viewport; the rest is clipped off and not displayed.

Figure 3.2. A world window and a viewport.

We want to describe not only how to do it in OpenGL, which is very easy, but also how it is done, to give
insight into the low-level algorithms used. We will work with only a 2D version here, but will later see how
these ideas extend naturally to 3D worlds viewed with a camera.

3.2.1. The mapping from the window to the viewport.

Figure 3.3 shows a world window and viewport in more detail. The world window is described by its left, top,
right, and bottom borders as W.l, W.t, W.r, and W.b, respectively2. The viewport is described likewise in the
coordinate system of the screen window (opened at some place on the screen), by V.l, V.t, V.r, and V.b, which
are measured in pixels.

2For

the sake of brevity we use l for left, t for top, etc. in mathematical formulas.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 3

Figure 3.3. Specifying the window and viewport.

The world window can be of any size and shape and in any position, as long as it is an aligned rectangle.
Similarly, the viewport can be any aligned rectangle, although it is of course usually chosen to lie entirely
within the screen window. Further, the world window and viewport dont have to have the same aspect ratio,
although distortion results if their aspect ratios differ. As suggested in Figure 3.4, distortion occurs because the
figure in the window must be stretched to fit in the viewport. We shall see later how to set up a viewport with an
aspect ratio that always matches that of the window, even when the user resizes the screen window.
y

W.l

V.r

V.l

W.t

V.t
W.r
x

W.b

V.b

window

viewport
sx
graphics
window

Figure 3.4. A picture mapped from a window to a viewport. Here some distortion is produced.
Given a description of the window and viewport, we derive a mapping or transformation, called the windowto-viewport mapping. This mapping is based on a formula that produces a point (sx, sy) in the screen window
coordinates for any given point (x, y) in the world. We want it to be a proportional mapping, in the sense that
if x is, say, 40% of the way over from the left edge of the window, then sx is 40% of the way over from the left
edge of the viewport. Similarly if y is some fraction, f, of the window height from the bottom, sy must be the
same fraction f up from the bottom of the viewport.
Proportionality forces the mappings to have a linear form:
sx = A * x + C
sy = B * y + D

(3.2)

for some constants A, B, C and D. The constants A and B scale the x and y coordinates, and C and D shift (or
translate) them.
How can A, B, C, and D be determined? Consider first the mapping for x. As shown in Figure 3.5,
proportionality dictates that (sx - V.l) is the same fraction of the total (V.r - V.l) as (x - W.l) is of the total (W.r W.l), so that

W.l

Computer Graphics

V.l

W.r

Chap 3

09/21/99

V.r

5:38 PM

page 4

Figure 3.5. Proportionality in mapping x to sx.

sx V.l
x W.l
V.r V .l = W.r W.l
or

sx =

V. r V . l
V. r V. l
x + (V . l
W . l)
W.r W.l
W. r W . l

Now identifying A as the part that multiplies x and C as the constant part, we obtain:

V. r V.l
, C = V . l A W. l
W . r W. l

Similarly, proportionality in y dictates that

sy V .b
y W.b
V.t V.b = W.t W.b
and writing sy as B y + D yields:

V.t V. b
, D = V. b B W. b
W . t W. b

Summarizing, the window to viewport transformation is:

sx = A x + C, sy = B y + D
with

(3.3)

V. r V.l
, C = V. l A W. l
W . r W. l
V.t V. b
, D = V. b B W. b
B=
W . t W. b
A=

The mapping can be used with any point (x, y) inside or outside the window. Points inside the window map to
points inside the viewport, and points outside the window map to points outside the viewport.
(Important!) Carefully check the following properties of this mapping using Equation 3.3:
a). if x is at the windows left edge: x = W.l, then sx is at the viewports left edge: sx = V.l.
b). if x is at the windows right edge then sx is at the viewports right edge.
c). if x is fraction f of the way across the window, then sx is fraction f of the way across the viewport.
d). if x is outside the window to the left, (x < w.l), then sx is outside the viewport to the left (sx < V.l), and
similarly if x is outside to the right.
Also check similar properties for the mapping from y to sy.
Example 3.2.1: Consider the window and viewport of Figure 3.6. The window has (W.l, W.r, W.b, W.t) = (0,
2.0, 0, 1.0) and the viewport has (V.l, V.r, V.b, V.t) = (40, 400, 60, 300).

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 5

Figure 3.6. An example of a window and viewport.

Using the formulas in Equation 3.2.2 we obtain
A = 180, C = 40,
B = 240, D = 60
Thus for this example, the window to viewport mapping is:
sx = 180 x + 40
sy = 240 y + 60
Check that this mapping properly maps various points of interest, such as:
Each corner of the window is indeed mapped to the corresponding corner of the viewport. For example, (2.0,
1.0) maps to (400, 300).
The center of the window (1.0, 0.5) maps to the center of the viewport (220, 180 ).
Practice Exercise 3.2.1. Building the mapping. Find values of A, B, C, and D for the case of a world window
10.0, 10.0, -6.0, 6.0) and a viewport (0, 600, 0, 400).
Doing it in OpenGL.
OpenGL makes it very easy to use the window to viewport mapping: it automatically passes each vertex it is
given (via a glVertex2*() command) through a sequence of transformations that carry out the desired
mapping. It also automatically clips off parts of objects lying outside the world window. All we need do is to set
up these transformations properly, and OpenGL does the rest.
For 2D drawing the world window is set by the function gluOrtho2D(), and the viewport is set by the function
glViewport(). These functions have prototypes:
void gluOrtho2D(GLdouble left, GLdouble right, GLdouble bottom, GLdouble
top);
which sets the window to have lower left corner (left, bottom) and upper right corner (right, top), and
void glViewport(GLint x, GLint y, GLint width, GLint height);
which sets the viewport to have lower left corner (x, y) and upper right corner (x + width, y + height).
By default the viewport is the entire screen window: if W and H are the width and height of the screen window,
respectively, the default viewport has lower left corner at (0, 0) and upper right corner at (W, H).

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 6

Because OpenGL uses matrices to set up all its transformations, gluOrtho2D()3 must be preceded by two set
up functions glMatrixMode(GL_PROJECTION) and glLoadIdentity(). (We discuss what is going on
behind the scenes here more fully in Chapter 5.)
Thus to establish the window and viewport used in Example 3.2.1 we would use:
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(0.0, 2.0, 0.0, 1.0);
glViewport(40, 60, 360, 240);

// sets the window

// sets the viewport

Hereafter every point (x, y) sent to OpenGL using glVertex2*(x, y) undergoes the mapping of Equation 3.3, and
edges are automatically clipped at the window boundary. (In Chapter 7 we see the details of how this is done in
3D, where it also becomes clear how the 2D version is simply a special case of the 3D version.)
It will make programs more readable if we encapsulate the commands that set the window into a function
setWindow() as shown in Figure 3.7. We also show setViewport() that hides the OpenGL details of
glViewport(..). To make it easier to use, its parameters are slightly rearranged to match those of
setWindow(), so they are both in the order left, right, bottom, top.
Note that for convenience we use simply the type float for the parameters to setWindow(). The parameters left,
right, etc. are automatically cast to type Gldouble when they are passed to gluOrtho2D(), as specified by
this function's prototype. Similarly we use the type int for the parameters to setViewport(), knowing the
arguments to glViewport() will be properly cast.
//--------------- setWindow --------------------void setWindow(float left, float right, float bottom, float top)
{
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(left, right, bottom, top);
}
//---------------- setViewport -----------------void setViewport(float left, float right, float bottom, float top)
{
glViewport(left, bottom, right left, top - bottom);
}
Figure 3.7. Handy functions to set the window and viewport.
It is worthwhile to look back and see what we used for a window and viewport in the early OpenGL programs
given in Chapter 2. In Figures 2.10 and 2.17 the programs used:
1). in main():
glutInitWindowSize(640,480);

// set screen window size

which set the size of the screen window to 640 by 480. The default viewport was used since no
glViewport() command was issued; the default viewport is the entire screen window.
2). in myInit():
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(0.0, 640.0, 0.0, 480.0);

The root ortho appears because setting the window this way is actually setting up a so-called
orthographic projection in 3D, as well see in Chapter 7.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 7

This set the world window to the aligned rectangle with corners (0,0) and (640.0, 480.0), just matching the
viewport size. So the underlying window to viewport mapping didnt alter anything. This was a reasonable
first choice for getting started.
Example 3.2.2: Plotting the sinc function revisited.
Putting these ingredients together, we can see what it takes to plot the sinc() function shape of Figure 3.1. With
OpenGL it is just a matter of defining the window and viewport. Figure 3.8 shows the required code, assuming
we want to plot the function from closely spaced x-values between 4.0 and 4.0, into a viewport with width 640
and height 480. (The window is set to be a little wider than the plot range to leave some cosmetic space around
the plot.)
void myDisplay(void) // plot the sinc function, using world coordinates
{
setWindow(-5.0, 5.0, -0.3, 1.0);
// set the window
setViewport(0, 640, 0, 480);
// set the viewport
glBegin(GL_LINE_STRIP);
for(GLfloat x = -4.0; x < 4.0; x += 0.1)
// draw the plot
glVertex2f(x, sin(3.14159 * x) / (3.14159 * x));
glEnd();
glFlush();
}
Figure 3.8. Plotting the sinc function.
Example 3.2.3: Drawing polylines from a file.
In Chapter 2 we drew the dinosaur shown in Figure 3.9 using the routine drawPolylineFile(
dino.dat) of Figure 2.22. The polyline data for the figure was stored in a file dino.dat. The world
window and viewport had not yet been introduced, so we just took certain things on faith or by default, and
luckily still got a picture of the dinosaur.

Figure 3.9. The dinosaur inside its world window.

Now we can see why it worked: the world window we used happened to enclose the data for the dinosaur (see
Case Study 2.4): All of the polylines in dino.dat lie inside a rectangle with corners (0, 0) and (640, 480), so
none are clipped with this choice of a window.
Armed with tools for setting the window and viewport, we can take more control of the situation. The next two
examples illustrate this.
Example 3.2.4. Tiling the screen window with the dinosaur motif.
To add some interest, we can draw a number of copies of the dinosaur in some pattern. If we lay them side by
side to cover the entire screen window its called tiling the screen window. The picture that is copied at
different positions is often called a motif. Tiling a screen window is easily achieved by using a different
viewport for each instance of the motif. Figure 3.10a shows a tiling involving 25 copies of the motif. It was
generated using:
a).
b).

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 8

Figure 3. 10. Tiling the display with copies of the dinosaur.

setWindow(0, 640.0, 0, 480.0);
//
for(int i = 0; i < 5; i++)
//
for(int j = 0; j < 5; j++)
//
{
glViewport(i * 64, j * 44, 64, 44); //
drawPolylineFile(dino.dat);
//
}

set a fixed window

for each column
for each row
set the next viewport
draw it again

(Its easier to use glViewport() here than setViewport(). What would the arguments to setViewport()
be if we chose to use it instead?) Each copy is drawn in a viewport 64 by 48 pixels in size, whose aspect ratio
64/48 matches that of the world window. This draws each dinosaur without any distortion.
Figure 3.10b shows another tiling, but here alternate motifs are flipped upside down to produce an intriguing effect. This was
done by flipping the window upside down every other iteration: interchanging the top and bottom values in
setWindow()4. (Check that this flip of the window properly affects B and D in the window to viewport transformation of
Equation 3.3 to flip the picture in the viewport.) Then the preceding double loop was changed to:
for(int i = 0; i < 5; i++)
for(int j = 0; j < 5; j++)
{
if((i + j) % 2 == 0)
setWindow(0.0, 640.0, 0.0, 480.0);
else
setWindow(0.0, 640.0, 480.0, 0.0);
glViewport(i * 64, j * 44, 64, 44);
drawPolylineFile(dino.dat);
}

// if (i + j) is even
// right side up window
// upside down window
// set the next viewport
// draw it again

Example 3.2.5. Clipping parts of a figure.

A picture can also be clipped by proper setting of the window. OpenGL automatically clips off parts of objects
that lie outside the world window. Figure 3.11a shows a figure consisting of a collection of hexagons of different
sizes, each slightly rotated relative to its neighbor. Suppose it is drawn by executing some function hexSwirl().
(We see how to write hexSwirl() in Section 3.6.) Also shown in part a are two boxes that indicate different
choices of a window. Parts b and c show what is drawn if these boxes are used for the world windows. It is
important to keep in mind that the same entire object is drawn in each case, using the code:

It might seem easier to invert the viewport, but OpenGL does not permit a viewport to have a negative
height.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 9

Figure 3.11. Using the window to clip parts of a figure.

setWindow(); // the window is changed for each picture
setViewport(); // use the same viewport for each picture
hexSwirl(); // the same function is called
What is displayed, on the other hand, depends on the setting of the window.
Zooming and roaming.
The example in Figure 3.11 points out how changing the window can produce useful effects. Making the window
smaller is much like zooming in on the object with a camera. Whatever is in the window must be stretched to fit
in the fixed viewport, so when the window is made smaller there must be greater enlargement of the portion
inside. Similarly making the window larger is equivalent to zooming out from the object. (Visualize how the
dinosaur would appear if the window were enlarged to twice the size it has in Figure 3.9.) A camera can also
roam (sometimes called pan) around a scene, taking in different parts of it at different times. This is easily
accomplished by shifting the window to a new position.
Example 3.2.6. Zooming in on a figure in an animation.
Consider putting together an animation where the camera zooms in on some portion of the hexagons in figure
3.11. We make a series of pictures, often called frames, using a slightly smaller window for each one. When the
frames are displayed in rapid succession the visual effect is of the camera zooming in on the object.
Figure 3.12 shows a few of the windows used: they are concentric and have a fixed aspect ratio, but their size
diminishes for each successive frame. Visualize what is drawn in the viewport for each of these windows.

Figure 3.12. Zooming in on the swirl of hexagons. (file: fig3.12.bmp)

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 10

A skeleton of the code to achieve this is shown in Figure 3.13. For each new frame the screen is cleared, the
window is made smaller (about a fixed center, and with a fixed aspect ratio), and the figure within the window is
drawn in a fixed viewport.
float cx = 0.3, cy = 0.2; //center of the window
float H, W = 1.2, aspect = 0.7; // window properties
set the viewport
for(int frame = 0; frame < NumFrames; frame++) // for each frame
{
clear the screen
// erase the previous figure
W *= 0.7;
// reduce the window width
H = W * aspect;
// maintain the same aspect ratio
setWindow(cx - W, cx + W, cy - H, cy + H); //set the next window
hexSwirl();
// draw the object
}
Figure 3.13. Making an animation.
Achieving a Smooth Animation.
The previous approach isnt completely satisfying, because of the time it takes to draw each new figure. What the
user sees is a repetitive cycle of:
a). Instantaneous erasure of the current figure;
b). A (possibly) slow redraw of the new figure.
The problem is that the user sees the line-by-line creation of the new frame, which can be distracting. What the
user would like to see is a repetitive cycle of:
a). A steady display of the current figure;
b). Instantaneous replacement of the current figure by the finished new figure;
The trick is to draw the new figure somewhere else while the user stares at the current figure, and then to
move the completed new figure instantaneously onto the users display. OpenGL offers double-buffering
to accomplish this. Memory is set aside for an extra screen window which is not visible on the actual display,
and all drawing is done to this buffer. (The use of such off-screen memory is discussed fully in Chapter 10.)
The command glutSwapBuffers() then causes the image in this buffer to be transferred onto the screen
window visible to the user.
To make OpenGL reserve a separate buffer for this, use GLUT_DOUBLE rather than GLUT_SINGLE in the
routine used in main() to initialize the display mode:
glutInitDisplayMode(GLUT_DOUBLE | GLUT_RGB); // use double buffering
The command glutSwapBuffers() would be placed directly after drawPolylineFile() in the code of
Figure 3.13. Then, even if it takes a substantial period for the polyline to be drawn, at least the image will
change abruptly from one figure to the next in the animation, producing a much smoother and visually
comfortable effect.
Practice Exercise 3.2.2. Whirling swirls. As another example of clipping and tiling, Figure 3.14a shows the
swirl of hexagons with a particular window defined. The window is kept fixed in this example, but the viewport
varies with each drawing. Figure 3.14b shows a number of copies of this figure laid side by side to tile the
display. Try to pick out the individual swirls. (Some of the swirls have been flipped: which ones?) The result is
dazzling to the eye, in part due to the eyes yearning to synthesize many small elements into an overall pattern.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 11

Figure 3.14. a). Whirling hexagons in a fixed window. b). A tiling formed using many viewports.
Except for the flipping, the code shown next creates this pattern. Function myDisplay() sets the window once,
then draws the clipped swirl again and again in different viewports.
void myDisplay(void)
{
clear the screen
setWindow(-0.6, 0.6, -0.6, 0.6); // the portion of the swirl to draw
for(int i = 0; i < 5; i++)
// make a pattern of 5 by 4 copies
for(int j = 0; j < 4; j++)
{
int L = 80; // the amount to shift each viewport
setViewport(i * L, L + i * L, j * L, L + j * L); // the next viewport
hexSwirl();
}
}
Type this code into an OpenGL environment, and experiment with the figures it draws. Taking a cue from a
previous example, determine how to flip alternating figures upside down.

3.2.2. Setting the Window and Viewport Automatically.

We want to see how to choose the window and viewport in order to produce appropriate pictures of a scene. In
some cases the programmer (or possibly the user at run-time) can input the window and viewport specifications
to achieve a certain effect; in other cases one or both of them are set up automatically, according to some
requirement for the picture. We discuss a few alternatives here.
Setting of the Window.
Often the programmer does not know where or how big the object of interest lies in world coordinates. The
object might be stored in a file like the dinosaur earlier, or it might be generated procedurally by some
algorithm whose details are not known. In such cases it is convenient to let the application determine a good
window to use.
The usual approach is to find a window that includes the entire object: to achieve this the objects extent must
be found. The extent (or bounding box) of an object is the aligned rectangle that just covers it. Figure 3.15
shows a picture made up of several line segments. The extent of the figure, shown as a dashed line, is (left,
right, bottom, top) = (0.36, 3.44, -0.51, 1.75).

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 12

(0.36, 1.75)

extent

x
sy
(3.44, -0.51)

Figure 3.15. Using the Extent as the Window.

How can the extent be computed for a given object? If all the endpoints of its lines are stored in an array pt[i],
for i = 0, 2, ..., n -1 the extent can be computed by finding the extreme values of the x- and y- coordinates in this
array. For instance, the left side of the extent is the smallest of the values pt[i].x. Once the extent is known,
the window can be made identical to it.
If, on the other hand, an object is procedurally defined, there may be no way to determine its extent ahead of
time. In such a case the routine may have to be run twice:
Pass 1: Execute the drawing routine, but do no actual drawing; just compute the extent. Then set the window.
Pass 2: Execute the drawing routine again. Do the actual drawing.
Automatic setting of the viewport to Preserve Aspect Ratio.
Suppose you want to draw the largest undistorted version of a figure that will fit in the screen window. For this
you need to specify a viewport that has the same aspect ratio as the world window. A common wish is to find
the largest such viewport that will fit inside the screen window on the display.
Suppose the aspect ratio of the world window is know to be R, and the screen window has width W and height
H. There are two distinct situations: the world window may have a larger aspect ratio than the screen window (R
> W/H), or it may have a smaller aspect ratio (R < W/H). The two situations are shown in Figure 3.16.

Figure 3.16. Possible aspect ratios for the world and screen windows.
Case a): R > W/H. Here the world window is short and stout relative to the screen window, so the viewport with
a matching aspect ratio R will extend fully across the screen window, but will leave some unused space above
or below. At its largest, therefore, it will have width W and height W/R, so the viewport is set using (check that
this viewport does indeed have aspect ratio R):
setViewport(0, W, 0, W/R);
Case b): R < W/H. Here the world window is tall and narrow relative to the screen window, so the viewport of
matching aspect ratio R will reach from the top to the bottom of the screen window, but will leave some unused
space to the left or right. At its largest it will have height H but width HR, so the viewport is set using:

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 13

setViewport(0, H * R, 0, H);
Example 3.2.7: A tall window. Suppose the window has aspect ratio R = 1.6 and the screen window has H =
200 and W = 360, and hence W/H = 1.8. Therefore Case b) applies, and the viewport is set to have a height of
200 pixels and a width of 320 pixels.
Example 3.2.8: A short window. Suppose R = 2 and the screen window is the same as in the example above.
Then case a) applies, and the viewport is set to have a height of 180 pixels and a width of 360 pixels.
Resizing the screen window, and the resize event.
In a windows-based system the user can resize the screen window at run-time, typically by dragging one of its
corners with the mouse. This action generates a resize event that the system can respond to. There is a function
in the OpenGL utility toolkit, glutReshape() that specifies a function to be called whenever this event
occurs:
glutReshape(myReshape);

//specifies the function called on a resize event

(This statement appears in main() along with the other calls that specify callback functions.) The registered
function is also called when the window is first opened. It must have the prototype:
void myReshape(GLsizei W, GLsizei H);
When it is executed the system automatically passes it the new width and height of the screen window, which it
can use in its calculations. (GLsizei is a 32 bit integer see Figure 2.7.)
What should myReshape() do? If the user makes the screen window bigger the previous viewport could still be
used (why?), but it might be desired to increase the viewport to take advantage of the larger window size. If the
user makes the screen window smaller, crossing any of the boundaries of the viewport, you almost certainly want
to recompute a new viewport.
Making a matched viewport.
One common approach is to find a new viewport that a) fits in the new screen window, and b) has the same
aspect ratio as the world window. Matching the aspect ratios of the viewport and world window in this way
will prevent distortion in the new picture. Figure 3.17 shows a version of myReshape() that does this: it finds
the largest matching viewport (matching the aspect ratio, R, of the window), that will fit in the new screen
window. The routine obtains the (new) screen window width and height through its arguments. Its code is a
simple embodiment of the result in Figure 3.16.
void myReshape(GLsizei W, GLsizei H)
{
if(R > W/H) // use (global) window aspect ratio
setViewport(0, W, 0, W/R);
else
setViewport(0, H * R, 0, H);
}
Figure 3.17. Using a reshape function to set the largest matching viewport upon a resize event.
Practice Exercises.
3.2.3. Find the bounding box for a polyline. Write a routine that computes the extent of the polyline stored in the
array of points pt[i], for i = 0, 2, ..., n 1.
3.2.4. Matching the Viewport. Find the matching viewport for a window with aspect ratio .75 when the screen
window has width 640 and height 480.
3.2.5. Centering the viewport. (Dont skip this one!) Adjust the myReshape() routine above so that the
viewport, rather than lying in the lower left corner of the display, is centered both vertically and horizontally in
the screen window.
3.2.6. How to squash a house. Choose a window and a viewport so that a square is squashed to half its proper
height. What are the coefficients A, B, C, and D in this case?

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 14

3.2.7. Calculation of the mapping. Find the coefficients A, B, C, and D of the window to viewport mapping for
a window given by (-600, 235, -500, 125) and a viewport (20, 140, 30, 260). Does distortion occur for figures
drawn in the world? Change the right border of the viewport so that distortion will not occur.

3.3. Clipping Lines.

Clipping is a fundamental task in graphics, needed to keep those parts of an object that lie outside a given
region from being drawn. A large number of clipping algorithms have been developed. In an OpenGL
environment each object is automatically clipped to the world window using a particular algorithm
(which we examine in detail in Chapter 7 for both 2D and 3D objects.)
Because OpenGL clips for you there may be a temptation to skip a study of the clipping process. But the ideas
that are used to develop a clipper are basic and arise in diverse situations; we will see a variety of approaches to
clipping in later chapters. And its useful to know how to pull together a clipper as needed when a tool like
OpenGL is not being used.
We develop a clipping algorithm here that clips off outlying parts of each line segment presented to it. This
algorithm can be incorporated in a line-drawing routine if we do not have the benefit of the clipping performed
by OpenGL. An implementation of a class that draws clipped lines is developed in Case Study 3.3.

3.3.1. Clipping a Line.

In this section we describe a classic line-clipping algorithm, the Cohen-Sutherland clipper, that computes which
part (if any) of a line segment with endpoints p1 and p2 lies inside the world window, and reports back the
endpoints of that part.
Well develop the routine clipSegment(p1, p2, window) that takes two 2D points and an aligned
rectangle. It clips the line segment defined by endpoints p1 and p2 to the window boundaries. If any portion of
the line remains within the window, the new endpoints are placed in p1 and p2, and 1 is returned (indicating
some part of the segment is visible). If the line is completely clipped out, 0 is returned (no part is visible).
Figure 3.18 shows a typical situation covering some of the many possible actions for a clipper. clipSegment(
) does one of four things to each line segment:
A
B

w indow

C
E
D

Figure 3.18. Clipping Lines at window boundaries.

If the entire line lies within the window, (e.g. segment CD): it returns 1.
If the entire line lies outside the window, (e.g. segment AB): it returns 0.
If one endpoint is inside the window and one is outside (e.g. segment ED): the function clips the portion of the
segment that lies outside the window and returns 1.
If both endpoints are outside the window, but a portion of the segment passes through it, (e.g. segment AE): it
clips both ends and returns 1.
There are many possible arrangements of a segment with respect to the window. The segment can lie to the left,
to the right, above, or below the window; it can cut through any one (or two) window edges, and so on. We
therefore need an organized and efficient approach that identifies the prevailing situation and computes new
endpoints for the clipped segment. Efficiency is important because a typical picture contains thousands of line

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 15

segments, and each must be clipped against the window. The CohenSutherland algorithm provides a rapid
divide-and-conquer attack on the problem. Other clipping methods are discussed beginning in Chapter 4.

3.3.2. The Cohen-Sutherland Clipping Algorithm

The Cohen-Sutherland algorithm quickly detects and dispenses with two common cases, called trivial accept
and trivial reject. As shown in Figure 3.19, both endpoints of segment
C

w indow
B

D
A

Figure 3.19. Trivial acceptance or rejection of a line segment.

AB lie within window W, and so the whole segment AB must lie inside. Therefore AB can be trivially
accepted: it needs no clipping. This situation occurs frequently when a large window is used that encompasses
most of the line segments. On the other hand, both endpoints C and D lie entirely to one side of W, and so
segment CD must lie entirely outside. It is trivially rejected, and nothing is drawn. This situation arises
frequently when a small window is used with a dense picture that has many segments outside the window.
Testing for a trivial accept or trivial reject.
We want a fast way to detect whether a line segment can be trivially accepted or rejected. To facilitate this, an
inside-outside code word is computed for each endpoint of the segment. Figure 3.20 shows how it is done.
Point P is to the left and above the window W. These two facts are recorded in a code word for P: a T (for
TRUE) is seen in the field for is to the left of, and is above. An F (for FALSE) is seen in the other two
fields, is to the right of, and is below.
P
is P to t he left of W?
code for P: T T F F
w indow
is P above W?

is P to t he right of W?

Figure 3.20. Encoding how point P is disposed with respect to the window.
For example, if P is inside the window its code is FFFF; if P is below but neither to the left nor right its code is
FFFT. Figure 3.21 shows the nine different regions possible, each with its code.
TTFF

FTFF

FTTF

TFFF

FFFF

FFTF

window
TFFT

FFFT

FFTT

Figure 3.21. Inside-outside codes for a point.

We form a code word for each of the endpoints of the line segment being tested. The conditions of trivial accept
and reject are easily related to these code words:

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 16

Trivial accept: Both code words are FFFF;

Trivial reject: the code words have an F in the same position: both points are to the left of the window, or
both are above, etc.

The actual formation of the code words and tests can be implemented very efficiently using the bit
manipulation capabilities of C/ C++, as we describe in Case Study 3.3.
Chopping when there is neither trivial accept nor reject.
The Cohen-Sutherland algorithm uses a divide-and-conquer strategy. If the segment can neither be trivially
accepted nor rejected it is broken into two parts at one of the window boundaries. One part lies outside the
window and is discarded. The other part is potentially visible, so the entire process is repeated for this segment
against another of the four window boundaries. This gives rise to the strategy:
do{
form the code words for p1 and p2
if (trivial accept) return 1;
if (trivial reject) return 0;
chop the line at the next window border; discard the outside part;
} while(1);
The algorithm terminates after at most four times through the loop, since at each iteration we retain only the
portion of the segment that has survived testing against previous window boundaries, and there are only four
such boundaries. After at most four iterations trivial acceptance or rejection is assured.
How is the chopping at each boundary done? Figure 3.22 shows an example involving the right edge of the
window.
P1
w indow

t op

dely

P2
bott om

de lx

left

right

Figure 3.22. Clipping a segment against an edge.

Point A must be computed. Its x-coordinate is clearly W.right, the right edge position of the window. Its ycoordinate requires adjusting p1.y by the amount d shown in the figure. But by similar triangles

d
e
=
dely delx
where e is p1.x - W.right and:
(3.4)

delx = p2.x - p1.x;

dely = p2.y - p1.y;

are the differences between the coordinates of the two endpoints. Thus d is easily determined, and the new
p1.y is found by adding an increment to the old as
p1.y += (W.right - p1.x)

Computer Graphics

Chap 3

(3.5)

* dely / delx

09/21/99

5:38 PM

page 17

Similar reasoning is used for clipping against the other three edges of window.
In some of the calculations the term dely/delx occurs, and in others it is delx/dely. One must always be
concerned about dividing by zero, and in fact delx is zero for a vertical line, and dely is 0 for a horizontal
line. But as discussed in the exercises the perilous lines of code are never executed when a denominator is zero,
so division by zero will not occur.
These ideas are collected in the routine clipSegment( ) shown in Figure 3.23. The endpoints of the segment
are passed by reference, since changes made to the endpoints by clipSegment() must be visible in the
calling routine. (The type Point2 holds a 2D point, and the type RealRect holds an aligned rectangle. Both
types are described fully in Section 3.4.)
int clipSegment(Point2& p1, Point2& p2, RealRect W)
{
do{
if(trivial accept) return 1; // some portion survives
if(trivial reject) return 0; // no portion survives
if(p1 is outside)
{
if(p1 is to the left) chop against the left edge
else if(p1 is to the right) chop against the right edge
else if(p1 is below) chop against the bottom edge
else if(p1 is above) chop against the top edge
}
else
// p2 is outside
{
if(p2 is to the left)chop against the left edge
else if(p2 is to the right)chop against the right edge
else if(p2 is below) chop against the bottom edge
else if(p2 is above)chop against the top edge
}
}while(1);
}
Figure 3.23. The Cohen-Sutherland line clipper (pseudocode).
Each time through the do loop the code for each endpoint is recomputed and tested. When trivial acceptance
and rejection fail, the algorithm tests whether p1 is outside, and if so it clips that end of the segment to a
window boundary. If p1 is inside then p2 must be outside (why?) so p2 is clipped to a window boundary.
This version of the algorithm clips in the order left, then right, then bottom, and then top. The choice of order is
immaterial if segments are equally likely to lie anywhere in the world. A situation that requires all four clips is
shown in Figure 3.24. The first clip
p2
D
B

Figure 3.24. A segment that requires four clips.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 18

changes p1 to A ; the second alters p2 to B; the third finds p1 still outside and below and so changes A to C ;
and the last changes p2 to D. For any choice of ordering for the chopping tests, there will always be a situation
in which all four clips are necessary.
Clipping is a fundamental operation that has received a lot of attention over the years. Several other approaches
have been developed. We examine some of them in the Case Studies at the end of this chapter, and in Chapter
4.
3.3.2. Hand Simulation of clipSegment( ).
Go through the clipping routine by hand for the case of a window given by (left, right, bottom, top) = (30, 220,
50, 240) and the following line segments:
1). p1=(40,140), p2=(100,200);
2). p1=(10,270), p2=(300,0);
3). p1=(20,10), p2=(20,200);
4). p1=(0,0), p2=(250,250);
In each case determine the endpoints of the clipped segment, and for a visual check, sketch the situation on
graph paper.

3.4. Developing the Canvas Class.

One must not always think that feeling is everything.
Art is nothing without form.
Gustave Flaubert
There is significant freedom in working in world coordinates, and having primitives be clipped and properly
mapped from the window to the viewport. But this freedom must be managed properly. There are so many
interacting ingredients (points, rectangles, mappings, etc.) in the soup now we should encapsulate them and
restrict how the application programmer accesses them, to avoid subtle bugs. We should also insure that the
various ingredients are properly initialized.
It is natural to use classes and the data hiding they offer. So we develop a class called Canvas that provides a
handy drawing canvas on which to draw the lines, polygons, etc. of interest. It provides simple methods to
create the desired screen window and to establish a world window and viewport, and it insures that the window
to viewport mapping is well defined. It also offers the routines moveTo() and lineTo() that many
programmers find congenial, as well as the useful turtle graphics routines we develop later in the chapter.
There are many ways to define the Canvas class: the choice presented here should be considered only as a
starting point for your own version. We implement the class in this section using OpenGL, exploiting all of the
operations OpenGL does automatically (such as clipping). But in Case Study 3.4 we describe an entirely
different implementation (based on Turbo C++ in a DOS environment), for which we have to supply all of the
tools. In particular an implementation of the Cohen Sutherland clipper is used.

3.4.1. Some useful Supporting Classes.

It will be convenient to have some common data types available for use with Canvas and other classes. We
define them here as classes5, and show simple constructors and other functions for handling objects of each
type. Some of the classes also have a draw function to make it easy to draw instances of the class. Other
member functions (methods) will be added later as the need arises. Some of the methods are implemented
directly in the class definitions; the implementation of others is requested in the exercises, and only the
declaration of the method is given.

class Point2: A point having real coordinates.

The first supporting class embodies a single point expressed with floating point coordinates. It is shown with two
constructors, the function set() to set the coordinate values, and two functions to retrieve the individual
coordinate values.
class Point2
{
5

Students preferring to write in C can define similar types using struct's.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 19

public:
Point2() {x = y = 0.0f;}
// constructor1
Point2(float xx, float yy) {x = xx; y = yy;} // constructor2
void set(float xx, float yy) {x = xx; y = yy;}
float getX() {return x;}
float getY() {return y;}
void draw(void) { glBegin(GL_POINTS); // draw this point
glVertex2f((Glfloat)x, (Glfloat)y);
glEnd();}
private:
float x, y;
};
Note that values of x and y are cast to the type Glfloat when glVertex2f() is called. This is mot likely
unnecessary since the type Glfloat is defined on most systems as float anyway.

class IntRect: An aligned rectangle with integer coordinates.

To describe a viewport we need an aligned rectangle having integer coordinates. The class IntRect provides
this.
class IntRect
{
public:
IntRect() {l = 0; r = 100; b = 0; t = 100;}// constructors
IntRect(int left, int right, int bottom, int top)
{l = left; r = right; b = bottom; t = top;}
void set(int left, int right, int bottom, int top)
{l = left; r = right; b = bottom; t = top;}
void draw(void); // draw this rectangle using OpenGL
private:
int l, r, b, t;
};

class RealRect: An aligned rectangle with real coordinates.

A world window requires the use of an aligned rectangle having real values for its boundary position. (This
class is so similar to IntRect some programmers would use templates to define a class that could hold either
integer or real coordinates.)
class RealRect
{
same as intRect except use float instead of int
};
Practice Exercise 3.4.1. Implementing the classes. Flesh out these classes by adding other functions you think
would be useful, and by implementing the functions, such as draw() for intRect, that have only been
declared above.

3.4.2. Declaration of Class Canvas.

We declare the interface for Canvas in Canvas.h as shown in Figure 3.25. Its data members include the
current position, a window, a viewport, and the window to viewport mapping.
class Canvas {
public:
Canvas(int width, int height, char* windowTitle); // constructor
void setWindow(float l, float r, float b, float t);
void setViewport(int l, int r, int b, int t);
IntRect getViewport(void); // divulge the viewport data
RealRect getWindow(void); // divulge the window data

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 20

float getWindowAspectRatio(void);
void clearScreen();
void setBackgroundColor(float r, float g, float b);
void setColor(float r, float g, float b);
void lineTo(float x, float y);
void lineTo(Point2 p);
void moveTo(float x, float y);
void moveTo(Point2 p);
others later
private:
Point2 CP;
// current position in the world
IntRect viewport; // the current window
RealRect window; // the current viewport
others later
};
Figure 3.25. The header file Canvas.h.
The Canvas constructor takes the width and height of the screen window along with the title string for the
window. As we show below it creates the screen window desired, performing all of the appropriate
initializations. Canvas also includes functions to set and return the dimensions of the window and the viewport,
and to control the drawing and background color. (There is no explicit mention of data for the window to
viewport mapping in this version, as this mapping is managed silently by OpenGL. In Case Study 3.4 we add
members to hold the mapping for an environment that requires it.). Other functions shown are versions of
lineTo() and moveTo() that do the actual drawing (in world coordinates, of course). We add relative drawing
tools in the next section.
Figure 3.26 shows how the Canvas class might typically be used in an application. A single global object cvs is
created, which initializes and opens the desired screen window. It is made global so that callback functions such
as display() can see it. (We cannot pass cvs as a parameter to such functions, as their prototypes are fixed
by the rules of the OpenGL utility toolkit.) The display() function here sets the window and viewport, and
then draws a line, using Canvas member functions. Then a rectangle is created and drawn using its own member
function.
Canvas cvs(640, 480, try out Canvas);
// create a global canvas object
//<<<<<<<<<<<<<<<<<<<<<<<<<<<<< display >>>>>>>>>>>>>>>>>>>>>>
void display(void)
{
cvs.clearScreen();
// clear screen
cvs.setWindow(-10.0, 10.0, -10.0, 10.0);
cvs.setViewport(10, 460, 10, 460);
cvs.moveTo(0, -10.0); // draw a line
cvs.lineTo(0, 10.0);
RealRect box( -2.0, 2.0, -1.0, 1.0); // construct a box
box.draw();
// draw the box
. . .
}
//<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
void main(void)
{
// the window has already been opened in the Canvas constructor
cvs.setBackgroundColor(1.0, 1.0, 1.0); // background is white
cvs.setColor(0.0, 0.0, 0.0); // set drawing color
glutDisplayFunc(display);
glutMainLoop();
}
Figure 3.26. Typical usage of the Canvas class.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 21

The main() routine doesnt do any initialization: this has all been done in the Canvas constructor. The routine
main() simply sets the drawing and background colors, registers function display(), and enters the main
event loop. (Could these OpenGL-specific functions also be buried in Canvas member functions?) Note that
this application makes almost no OpenGL-specific calls, so it could easily be ported to another environment
(which used a different implementation of Canvas, of course).

3.4.3. Implementation of Class Canvas.

We show next some details of an implementation of this class when OpenGL is available. (Case Study 3.4
discusses an alternate implementation.) The constructor, shown in Figure 3.27, passes the desired width and
height (in pixels) to glutInitWindowSize(), and the desired title string to glutCreateWindow(). Some
fussing must be done to pass glutInit() the arguments it needs, even though they arent used here. (Normally
main() passes glutInit() the command line arguments, as we saw earlier. This cant be done here since we
will use a global Canvas object , cvs, which is constructed before main() is called.)
//<<<<<<<<<<<<<<<<<<<<< Canvas constructor >>>>>>>>>>>>>>>>
Canvas:: Canvas(int width, int height, char* windowTitle)
{
char* argv[1];
// dummy argument list for glutInit()
char dummyString[8];
argv[0] = dummyString; // hook up the pointer
int argc = 1;
// to satisfy glutInit()
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB);
glutInitWindowSize(width, height);
glutInitWindowPosition(20, 20);
glutCreateWindow(windowTitle); // open the screen window
setWindow(0, (float)width, 0, (float)height); //default world window
setViewport(0, width, 0, height); // default viewport
CP.set(0.0f, 0.0f);
// initialize the CP to (0, 0)
}
Figure 3.27. The constructor for Canvas OpenGL version.
Figure 3.28 shows the implementation of some of the remaining Canvas member functions. (Others are
requested in the exercises.) Function moveTo() simply updates the current position; lineTo() sends the CP as
the first vertex, and the new point (x, y) as the second vertex. Note that we dont need to use the window to
viewport mapping explicitly here, since OpenGL automatically applies it. The function setWindow() passes its
arguments to gluOrtho2D() after properly casting their types and loads them into Canvass window.
//<<<<<<<<<<<<<<<<<<<<<<<<<< moveTo >>>>>>>>>>>>>>>>>>
void Canvas:: moveTo(float x, float y)
{
CP.set(x, y);
}
//<<<<<<<<<<<<<<<<<<<<<<< lineTo >>>>>>>>>>>>>>>>>>>>>>>
void Canvas:: lineTo(float x, float y)
{
glBegin(GL_LINES);
glVertex2f((GLfloat)CP.x, (GLfloat)CP.y);
glVertex2f((GLfloat)x, (GLfloat)y);
// draw the line
glEnd();
CP.set(x, y); // update the CP
glFlush();
}
//<<<<<<<<<<<<<<<<<<<<<<<< set Window >>>>>>>>>>>>>>>>>>>>
void Canvas:: setWindow(float l, float r, float b, float t)
{
glMatrixMode(GL_PROJECTION);
glLoadIdentity();

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 22

gluOrtho2D((GLdouble)l, (GLdouble)r, (GLdouble)b, (GLdouble)t);

window.set(l, r, b, t);
}
Figure 3.28. Implementation of some Canvas member functions.
Practice Exercises.
3.4.2. Flesh out each of the member functions:
a). void setViewport(int l, int r, int b, int t);
b). IntRect getViewport(void);
c). RealRect getWindow(void);
d). void clearScreen(void);
e). void setBackgroundColor(float r, float g, float b);
f). void setColor(float r, float g, float b);
g). void lineTo(Point2 p);
h). void moveTo(Point2 p);
i). float getWindowAspectRatio(void)
3.4.3. Using Canvas for a simulation: Fibonacci numbers. The growth in the size of a rabbit population is
said to be modeled by the following equation [gardner61]:
yk = yk-1 + yk-2
where yk is the number of bunnies at the k-th generation. This model says that the number in this generation is
the sum of the numbers in the previous two generations. The initial populations are y0 = 1 and y1 = 1. Successive
values of yk are formed by substituting earlier values, and the resulting sequence is the well-known Fibonacci
sequence; 1, 1, 2, 3, 5, 8, 13. . . . A plot of the sequence yk versus k reveals the nature of this growth pattern.
Use the Canvas class to write a program that draws such a plot for a sequence of length N. Adjust the size of
the plot appropriately for different N. (The sequence grows very rapidly, so you may instead wish to plot the
logarithm of yk versus k.) Also plot the sequence of ratios pk = yk / yk-1 and watch how quickly this ratio
converges to the golden ratio.
3.4.4. Another Simulation: sinusoidal sequences. The following difference equation generates a sinusoidal
sequence:
yk = a yk-1 - yk-2

for k = 1, 2, . . .

where a is a constant between 0 and 2; yk is 0 for k < 0; and y0 = 1 (see [oppenheim83]). In general, one cycle
consists of S points if we set a = 2 cos(2/S). A good picture results with S = 40. Write a routine that draws
sequences generated in this fashion, and test it for various values of S.

3.5. Relative Drawing.

If we add just a few more drawing tools to our tool bag (which is the emerging class Canvas) certain drawing
tasks become much simpler. It is often convenient to have drawing take place at the current position (CP), and
to describe positions relative to the CP. We develop functions, therefore, whose parameters specify changes in
position: the programmer specifies how far to go along each coordinate to the next desired point.

3.5.1. Developing moveRel() and lineRel().

Two new routines are moveRel() and lineRel(). The function moveRel() is easy: it just moves the CP
through the displacement (dx, dy). The function lineRel(float dx, float dy) does this too, but it
first draws a line from the old CP to the new one. Both functions are shown in Figure 3.29.
void Canvas :: moveRel(float dx, float dy)
{
CP.set(CP.x + dx, CP.y + dy);

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 23

}
void Canvas :: lineRel(float dx, float dy)
{
float x = CP.x + dx, y = CP.y + dy;
lineTo(x, y);
CP.set(x, y);
}
Figure 3.29. The functions moveRel() and lineRel().
Example 3.5.1. An arrow marker. Markers of different shapes can be placed at various points in a drawing to
add emphasis. Figure 3.30 shows pentagram markers used to highlight the data points in a line graph.
y
4
3

4
6
3
5
Figure 3.30. Placing markers for emphasis.

Because the same figure is drawn at several different points it is convenient to be able to say simply
drawMarker() and have it be drawn at the CP. Then the line graph of Figure 3.30 can be drawn along with the
markers using code suggested by the pseudocode:
moveTo(first data point);
drawMarker();
// draw a marker there
for(each remaining data point)
{
lineTo(the next point); // draw the next line segment
drawMarker();
// draws it at the CP
}
Figure 3.31 shows an arrow-shaped marker, drawn using the routine in Figure 3.32. The arrow is positioned
with its uppermost point at the CP. For flexibility the arrow shape is parameterized through four size parameters
f, h, t, and w as shown. Function arrow() uses only lineRel(), and no reference is made to absolute
positions. Also note that although the CP is altered while drawing is going on, at the end the CP has been set
back to its initial position. Hence the routine produces no side effects (beyond the drawing itself).
worldCP
f

h
t
w
w
Figure 3.31. Model of an arrow.

void arrow(float f, float h, float t, float w)

{ // assumes global Canvas object: cvs
cvs.lineRel(-w - t / 2, -f);
// down the left side
cvs.lineRel(w, 0);

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 24

cvs.lineRel(0,
cvs.lineRel(t,
cvs.lineRel(0,
cvs.lineRel(w,
cvs.lineRel(-w

-h);
0);
h);
0);
- t / 2, f);

// across
// back up

}
Figure 3.32. Drawing an arrow using relative moves and draws.

3.5.2. Turtle Graphics.

The last tool we add for now is surprisingly convenient. It keeps track not only of where we are with the CP,
but also the direction in which we are headed. This is a form of turtlegraphics, which has been found to be a
natural way to program in graphics6. The notion is that a turtle, which is conceptually similar to the pen in a
pen plotter, migrates over the page, leaving a trail behind itself which appears as a line segment. The turtle is
positioned at the CP, headed in a certain direction called the current direction, CD. CD is the number of
degrees measured counterclockwise (CCW) from the positive x-axis.
It is easy to add functionality to the Canvas class to control the turtle. First, CD is added as a private data
member. Then we add three methods:
1). turnTo(float angle). Turn the turtle to the given angle, implemented as:
void Canvas:: turnTo(float angle) {CD = angle;}
2). turn(float angle). Turn the turtle through angle degrees counterclockwise:
void Canvas:: turn(angle){CD += angle;}
Use a negative argument to make a right turn. Note that a turn is a relative direction change: we dont specify
a direction, only a change in direction. This simple distinction provides enormous power in drawing complex
figures with the turtle.
3). forward(float dist, int isVisible). Move the turtle forward in a straight line from the CP
through a distance dist in the current direction CD, and update the CP. If isVisible is nonzero a visible
line is drawn; otherwise nothing is drawn.

Figure 3.33 shows that in going forward in direction CD the turtle just moves in x through the amount dist *
cos( * CD/180) and in y through the amount dist * sin( * CD/180), so the implementation of forward() is
immediate:
new worldCP

CD
old worldCP
Figure 3.33. Effect of the forward() routine.

void Canvas:: forward(float dist, int isVisible)

{
6Introduced by Seymour Papert at MIT as part of the LOGO language for teaching children how to program. See

e.g. [Abel81]

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 25

const float RadPerDeg = 0.017453393;

//radians per degree
float x = CP.x + dist * cos(RadPerDeg * CD);
float y = CP.y + dist * sin(RadPerDeg * CD);
if(isVisible)
lineTo(x, y);
else
moveTo(x, y);
}
Turtle graphics makes it easy to build complex figures out of simpler ones, as we see in the next examples.
Example 3.5.2. Building a figure upon a hook motif. The 3-segment hook motif shown in Figure 3.34a can
be drawn using the commands:
forward(3 * L, 1);
turn(90);
forward(L, 1);
turn(90);
forward(L, 1);
turn(90);

// L is the length of the short sides

for some choice of L. Suppose that procedure hook() encapsulates these instructions. Then the shape in Figure
3.34b is drawn using four repetitions of hook(). The figure can be positioned and oriented as desired by
choices of the initial CP and CD.
a).
b).

motif
Figure 3.34. Building a figure out of several turtle motions.
Example 3.5.3. Polyspirals. A large family of pleasing figures called polyspirals can be generated easily using
turtlegraphics. A polyspiral is a polyline where each successive segment is larger (or smaller) than its
predecessor by a fixed amount, and oriented at some fixed angle to the predecessor. A polyspiral is rendered by
the following pseudocode:
for(<some number of iterations>)
{
forward(length,1);
// draw a line in the current direction
turn(angle);
// turn through angle degrees
length += increment;
// increment the line length
}
Each time a line is drawn both its length and direction are incremented. If increment is 0, the figure neither
grows nor shrinks.. Figure 3.35 shows several polyspirals. The implementation of this routine is requested in the
exercises.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 26

Figure 3.35. Examples of polyspirals. Angles are: a). 60, b). 89.5, c). -144, d). 170.
Practice Exercises.
3.5.1. Drawing Turtle figures. Provide routines that use turtle motions to draw the three figures shown in
Figure 3.36. Can the turtle draw the shape in part c without lifting the pen and without drawing any line
twice?

a).

b).

c).

Figure 3.36. Other Simple Turtle Figures.

3.5.2. Drawing a well-known logo. Write a routine that makes a turtle draw the outline of the logo shown in
Figure 3.37. (It need not fill the polygons.)

Figure 3.37. A famous logo.

3.5.3. Driving the Turtle with Strings. We can use a shorthand notation to describe a figure. Suppose
F
means forward(d, 1);
{for some distance d}
L
means turn(60);
{left turn}
R
means turn(-60).
{right turn}
What does the following sequence of commands produce?
FLFLFLFRFLFLFLFRFLFLFLFR. (See Chapter 9 for a generalization of this that produces fractals!)
3.5.4. Drawing Meanders. A meander7 is a pattern like that in Figure 3.38a, often made up of a continuous
line meandering along some path. One frequently sees meanders on Greek vases, Chinese plates, or floor tilings
from various countries. The motif for the meander here is shown in Figure 3.38b. After each motif is drawn the
turtle is turned (how much?) to prepare it for drawing the next motif.

7Based on the name Maeander (which has modern name Menderes), a winding river in Turkey [Janson 86].

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 27

a).

b).

Figure 3.38. Example of a meander.

Write a routine that draws this motif, and a routine that draws this meander. (Meanders are most attractive if the
graphics package at hand supports the control of line thickness -- as OpenGL does -- so that forward() draws
thick lines.) A dazzling variety of more complex meanders can be designed, as suggested in later exercises. A
meander is a particular type of frieze pattern. Friezes are studied further in Chapter ???.
3.5.5. Other Classes of Meanders. Figure 3.39 shows two additional types of meanders. Write routines that
employ turtle graphics to draw them.

Figure 3.39. Additional figures for meanders.

3.5.6. Drawing Elaborate Meanders. Figure 3.40 shows a sequence of increasingly complex motifs for
meanders. Write routines that draw a meander for each of these motifs. What does the next most complicated
motif in this sequence look like, and what is the general principal behind constructing these motifs?

Figure 3.40. Hierarchy of meander motifs.

3.5.7. Implementing polyspiral. Write the routine polyspiral(float length, float angle,
float incr, int num) that draws a polyspiral consisting of num segments, the first having length
length. After each segment is drawn length is incremented by incr and the turtle turns through angle
angle.
3.5.8. Is a Polyspiral an IFS? Can a polyspiral be described in terms of an iterated function system as defined
in Chapter 2? Specify the function that is iterated by the turtle at each iteration.
3.5.9. Recursive form for Polyspiral(). Rewrite polyspiral()in a recursive form, so that
polyspiral() with argument dist calls polyspiral() with argument dist+inc. Put a suitable
stopping criterion in the routine.

3.6. Figures based on Regular Polygons.

To generalize is to be an idiot.
William Blake
Bees...by virtue of certain geometrical forethought...know that the hexagon is greater than the square and
triangle, and will hold more honey for the same expenditure of material.
Pappus of Alexandria
The regular polygons form a large and important family of shapes, often encountered in computer graphics. We
need efficient ways to draw them. In this section we examine how to do this, and how to create a number of
figures that are variations of the regular polygon.

3.6.1. The Regular Polygons.

First recall the definition of a regular polygon:

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 28

Definition: A polygon is regular if it is simple, if all its sides have equal lengths, and if adjacent sides meet at
equal interior angles.
As discussed in Chapter 1, a polygon is simple if no two of its edges cross each other (more precisely: only
adjacent edges can touch, and only at their shared endpoint). We give the name n-gon to a regular polygon
having n sides. Familiar examples are the 4-gon (a square), a 5-gon (a regular pentagon), 8-gon (a regular
octagon), and so on. A 3-gon is an equilateral triangle. Figure 3.41 shows various examples. If the number of
sides of an n-gon is large the polygon approximates a circle in appearance. In fact this is used later as one way
to implement the drawing of a circle.
n: 3

Figure 3.41. Examples of n-gons.

The vertices of an n-gon lie on a circle, the so-called parent circle of the n-gon, and their locations are easily
calculated. The case of the hexagon is shown in Figure 3.42 where the vertices lie equispaced every 60o around
the circle. The parent circle of radius R (not shown) is centered at the origin, and the first vertex P0 has been
placed on the positive x-axis. The other vertices follow accordingly, as Pi = ( R cos(i a), R sin(i a)), for i =
1,...,5, where a is 2/6 radians. Similarly, the vertices of the general n-gon lie at:
y
P2

P1 = (R cos(a), R sin(a))
a

R
P0

Figure 3.42. Finding the vertices of an 6-gon.

Pi = ( R cos( 2i / n ), R sin(2i / n )), for i = 0,..., n-1

(3.6)

Its easy to modify this n-gon. To center it at position (cx, cy) we need only add cx and cy to the x- and ycoordinates, respectively. To scale it by factor S we need only multiply R by S. To rotate through angle A we
need only add A to the arguments of cos() and sin(). More general methods for performing geometrical
transformations are discussed in Chapter 6.
It is simple to implement a routine that draws an n-gon, as shown in Figure 3.43. The n-gon is drawn centered at
(cx, cy), with radius radius, and is rotated through rotAngle degrees.
void ngon(int n, float cx, float cy, float radius, float rotAngle)
{
// assumes global Canvas object, cvs
if(n < 3) return;
// bad number of sides
double angle = rotAngle * 3.14159265 / 180; // initial angle
double angleInc = 2 * 3.14159265 /n;
//angle increment
cvs. moveTo(radius + cx, cy);
for(int k = 0; k < n; k++) // repeat n times
{
angle += angleInc;
cvs.lineTo(radius * cos(angle) + cx, radius * sin(angle) + cy);
}
}
Figure 3.43. Building an n-gon in memory.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 29

Example 3.6.1: A Turtle-driven n-gon. It is also simple to draw an n-gon using turtlegraphics. Figure 3.44
shows how to draw a regular hexagon. The initial position and direction of the turtle is indicated by the small
triangle. The turtle simply goes forward six times, making a CCW turn of 60 degrees between each move:

360 / n
R
L
Figure 3.44. Drawing a hexagon.
for (i = 0; i < 6; i++)
{
cvs.forward(L, 1);
cvs.turn(60);
}
One vertex is situated at the initial CP, and both CP and CD are left unchanged by the process. Drawing the
general n-gon, and some variations of it, is discussed in the exercises.

3.6.2. Variations on n-gons.

Interesting variations based on the vertices of an n-gon can also be drawn. The n-gon vertices may be connected
in various ways to produce a variety of figures, as suggested in Figure 3.45. The standard n-gon is drawn in
Figure 3.45a by connecting adjacent vertices, but Figure 3.45b shows a stellation (or star-like figure) formed by
connecting every other vertex. And Figure 3.45c shows the interesting rosette, formed by connecting each
vertex to every other vertex. We discuss the rosette next. Other figures are described in the exercises.
a).
b).
c).

Figure 3.45. A 7-gon and its offspring. a). the 7-gon, b). a stellation, c). a 7-rosette.
Example 3.6.2. The rosette, and the Golden 5-rosette.
The rosette is an n-gon with each vertex joined to every other vertex. Figure 3.46 shows 5-, 11-, and 17rosettes. A rosette is sometimes used as a test pattern for computer graphics devices. Its orderly shape readily
reveals any distortions, and the resolution of the device can be determined by noting the amount of crowding
and blurring exhibited by the bundle of lines that meet at each vertex.

Figure 3.46. The 5-, 11-, and 17-rosettes.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 30

Rosettes are easy to draw: simply connect every vertex to every other. In pseudocode this looks like
void Rosette(int N, float radius)
{
Point2 pt[big enough value for largest rosette];
generate the vertices pt[0],. . .,pt[N-1], as in Figure 3.43
for(int i = 0; i < N - 1; i++)
for(int j = i + 1; j < N ; j++)
{
cvs.moveTo(pt[i]); // connect all the vertices
cvs.lineTo(pt[j]);
}
}
The 5-rosette is particularly interesting because it embodies many instances of the golden ratio (recall Chapter
2). Figure 3.47a shows a 5-rosette, which is made up of an outer pentagon and an inner pentagram. The Greeks
saw a mystical significance in this figure. Its segments have an interesting relationship: Each segment is times
longer than the next smaller one (see the exercises). Also, because the edges of the star pentagram form an inner
pentagon, an infinite regression of pentagrams is possible, as shown in Figure 3.47b.
a).

b).
2

Figure 3.47. 5-rosette and Infinite regressions - pentagons and pentagrams.

Example 3.6.3. Figures based on two concentric n-gons.

Figures 3.48 shows some shapes built upon two concentric parent circles, the outer of radius R, and the inner of
radius fR for some fraction f. Each figure uses a variation of an n-gon whose radius alternates between the inner
and outer radii. Parts a) and b) show familiar company logos based on 6-gons and 10-gons. Part c) is based on
the 14-gon, and part d) shows the inner circle explicitly.
a).
b)
c)
d).

radius R

radius f R

Figure 3.48. A family of Famous Logos.

Practice Exercises.
3.6.1. Stellations and rosettes. The pentagram is drawn by connecting every other point as one traverses
around a 5-gon. Extend this to an arbitrary odd-valued n-gon and develop a routine that draws this so-called
stellated polygon. Can it be done with a single initial moveTo() followed only by lineTo()s (that is,
without lifting the pen)? What happens if n is even?
3.6.2. How Many Edges in an N-rosette? Show that a rosette based on an N-gon, an N-rosette, has N(N - 1) / 2
edges. This is the same as the number of clinks one hears when N people are seated around a table and
everybody clinks glasses with everyone else.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 31

3.6.3. Prime Rosettes. If a rosette has a prime number N of sides, it can be drawn without lifting the pen, that
is, by using only lineTo(). Start at vertex v0 and draw to each of the others in turn: v1,v2, v3, . . . until v0 is
again reached and the polygon is drawn. Then go around again drawing lines, but skip a vertex each time that
is, increment the index by 2 thereby drawing to v2, v4, . . . , v0. This will require going around twice to arrive
back at v0. (A modulo operation is performed on the indices so that their values remain between 0 and N-1.)
Then repeat this, incrementing by 3: v3, v6, v0, . . . , v0. Each repeat draws exactly N lines. Because there are N(N
- 1) / 2 lines in all, the process repeats (N - 1) / 2 times. Because the number of vertices is a prime, no pattern is
ever repeated until the drawing is complete. Develop and test a routine that draws prime rosettes in this way.
3.6.4. Rosettes with an odd number of sides. If n is prime we know the n-rosette can be drawn as a single
polyline without lifting the pen. It can also be drawn as a single polyline for any odd value of n. Devise a
method that does this.
3.6.5. The Geometry of the Star Pentagram. Show that the length of each segment in the 5-rosette stands in
the golden ratio to that of the next smaller one. One way to tackle this is to show that the triangles of the star
pentagram are golden triangles with an inner angle of / 5 radians. Show that 2 * cos( / 5) = and 2 *
cos(2 / 5) = 1 / . Another approach uses only two families of similar triangles in the pentagram and the
relation 3 = 2 + 1 satisfied by .
3.6.6. Erecting Triangles on n-gon legs. Write a routine that draws figures like the logo in part a of Figure 3.48
for any value of f, positive or negative. What is a reasonable geometric interpretation of negative f?
3.6.7. Drawing the Star with Relative Moves and Draws. Write a routine to draw a pentagram that uses only
relative moves and draws, centering the star at the CP.
3.6.8. Draw a pattern of stars. Write a routine to draw the pattern of 21 stars shown in Figure 3.49. The small
stars are positioned at the vertices of an n-gon.

Figure 3.49. A star pattern.

3.6.9. New points on the 7-gram. Figure 3.50 shows a figure formed from the 7 points of a 7-gon, centered
at the origin. The first point lies at (R, 0). Instead of connecting consecutive points around the 7-gon, two
intermediary points are skipped. (This is a form of stellation of an n-gon.) Find the coordinates of point P,
where two of the edges intersect.

P
(R, 0)

Figure 3.50. A 7-gram.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 32

3.6.10. Turtle drawings of the n-gon. Write turtleNgon(int numSides, float length) that uses
turtlegraphics to draw an n-gon with numSides sides and a side of length length.
3.6.11. Polygons sharing an edge. Write a routine that draws n-gons, for n = 3,, 12, on a common edge, as
in Figure 3.51.

Figure 3.51. N-gons sharing a common edge.

3.6.12. A more elaborate figure. Write a routine that draws the shape in Figure 3.52 by drawing repeated
hexagons rotated relative to one another.

Figure 3.52. Repeated use of turtle commands.

3.6.13. Drawing a famous logo. The esteemed logo shown in Figure 3.53 consists of three instances of a motif,
rotated a certain amount with respect to each other. Show a routine that draws this shape using turtlegraphics.

Figure 3.53. Logo of the University of Massachusetts.

3.6.14. Rotating Pentagons: animation. Figure 3.54 shows a pentagram oriented with some angle of rotation
within a pentagon, with corresponding vertices joined together. Write a program that animates this figure. The
configuration is drawn using some initial angle A of rotation for the pentagram. After a short pause it is erased
and then redrawn but with a slightly larger angle A. This process repeats until a key is pressed.

Figure 3.54. Rotating penta-things.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 33

3.7. Drawing Circles and Arcs.

Drawing a circle is equivalent to drawing an n-gon that has a large number of vertices. The n-gon resembles a
circle (unless it is scrutinized too closely). The routine drawCircle() shown in Figure 3.55 draws a 50-sided
n-gon, by simply passing its parameters on to ngon(). It would be more efficient to write drawCircle() from
scratch, basing it on the code of Figure 3.43.
void drawCircle(Point2 center, float radius)
{
const int numVerts = 50;
// use larger for a better circle
ngon(numVerts, center.getX(), center.getY(), radius, 0);
}
Figure 3.55. Drawing a circle based on an 50-gon.

3.7.1. Drawing Arcs.

Many figures in art, architecture, and science involve arcs of circles placed in pleasing or significant
arrangements. An arc is conveniently described by the position of the center, c,and radius, R, of its parent
circle, along with its beginning angle a and the angle b through which it sweeps. Figure 3.56 shows such an
arc. We assume that if b is positive the arc sweeps in a CCW direction from a. If b is negative it sweeps in a
CW fashion. A circle is a special case of an arc, with a sweep of 360o.
y

a
c
x

Figure 3.56. Defining an arc.

We want a routine, drawArc( ), that draws an arc of a circle. The function shown in Figure 3.57 approximates
the arc by part of an n-gon, using moveTo( ) and lineTo(). Successive points along the arc are found by
computing a cos() and sin() term each time through the main loop. If sweep is negative the angle automatically
decreases each time through.
void drawArc(Point2 center, float radius, float startAngle, float sweep)
{
// startAngle and sweep are in degrees
const int n = 30; // number of intermediate segments in arc
float angle = startAngle * 3.14159265 / 180; // initial angle in radians
float angleInc = sweep * 3.14159265 /(180 * n); // angle increment
float cx = center.getX(), cy = center.getY();
cvs.moveTo(cx + radius * cos(angle), cy + radius * sin(angle));
for(int k = 1; k < n; k++, angle += angleInc)
cvs.lineTo(cx + radius * cos(angle), cy + radius * sin(angle));
}
Figure 3.57. Drawing an arc of a circle.
The CP is left at the last point on the arc. (In some cases one may wish to omit the initial moveTo() to the first
point on the arc, so that the arc is connected to whatever shape was being drawn when drawArc() is called.)
A much faster arc drawing routine is developed in Chapter 5 that avoids the repetitive calculation of so many
sin() and cos() functions. It may be used freely in place of the procedure here.
With drawArc( ) in hand it is a simple matter to build the routine drawCircle(Point2 center,
float radius) that draws an entire circle (how?).

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 34

The routine drawCircle( ) is called by specifying a center and radius, but there are other ways to describe a
circle, which have important applications in interactive graphics and computer-aided design. Two familiar ones
are:
1). The center is given, along with a point on the circle. Here drawCircle( ) can be used as soon as the
radius is known. If c is the center and p is the given point on the circle, the radius is simply the distance from c
to p, found using the usual Pythagorean Theorem.
2). Three points are given through which the circle must pass. It is known that a unique circle passes through
any three points that don't lie in a straight line. Finding the center and radius of this circle is discussed in
Chapter 4.
Example 3.7.1. Blending Arcs together. More complex shapes can be obtained by using parts of two circles
that are tangent to one another. Figure 3.58 illustrates the underlying principle. The two circles are tangent at
point A, where they share tangent line L. Because of this the two arcs shown by the thick curve blend together
seamlessly at A with no visible break or corner. Similarly the arc of a circle blends smoothly with any tangent
line, as at point B.

Figure 3.58. Blending arcs using tangent circles.

Practice Exercises.
3.7.1. Circle Figures in Philosophy. In Chinese philosophy and religion the two principles of yin and yang
interact to influence all creatures destinies. Figure 3.59 shows the exquisite yinyang symbol. The dark portion,
yin, represents the feminine aspect, and the light portion, yang, represents the masculine. Describe in detail the
geometry of this symbol, supposing it is centered in some coordinate system.

use figure 4.3 from 1st edition

Figure 3.59. The yin-yang symbol.
3.7.2. The Seven Pennies. Describe the configuration shown in Figure 3.60 in which six pennies fit snugly
around a center penny. Use symmetry arguments to explain why the fit is exact; that is, why each of the outer
pennies exactly touches its three neighbors.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 35

Figure 3.60. The seven circles.

3.7.3. A famous logo. Figure 3.61 shows a well-known automobile logo. It is formed by erecting triangles
inside an equilateral triangle, but the outer triangle is replaced by two concentric circles. After determining the
proper positions for the three inner points, write a routine to draw this logo.

Figure 3.61. A famous logo.

3.7.4. Drawing clocks and such. Circles and lines may be made tangent in a variety of ways to create pleasing
smooth curves, as in Figure 3.62a. Figure 3.62b shows the underlying lines and circles. Write a routine that
draws this basic clock shape
a).
b).

Figure 3.62. Blending arcs to form smooth curves.

3.7.5. Drawing rounded rectangles. Figure 3.63 shows an aligned rectangle with rounded corners. The
rectangle has width W and aspect ratio R, and each corner is described by a quarter-circle of radius r = g W for
some fraction g. Write a routine drawRoundRect(float W, float R, float g) that draws this
rectangle centered at the CP. The CP should be left at the center when the routine exits.
r
RW

Figure 3.63. A rounded rectangle.

3.7.6. Shapes involving arcs. Figure 3.64 shows two interesting shapes that involve circles or arcs. One is
similar to the Atomic Energy Commission symbol (How does it differ form the standard symbol?). Write and
test two routines that draw these figures.
caution radiation

Figure 3.64. Shapes based on arcs.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 36

3.7.7. A tear drop. A tear drop shape that is used in many ornamental figures is shown in Figure 3.65a. As
shown in part b) it consists of a circle of given radius R snuggled down into an angle . What are the
coordinates of the circles center C for a given R and ? What are the initial angle of the arc, and its sweep?
Develop a routine to draw a tear drop at any position and in any orientation.
b).

a).

Figure 3.65. The tear drop and its construction.

3.7.8. Drawing Patterns of Tear Drops. Figure 3.66 show some uses of the tear drop. Write a routine that
draws each of them.
a).

b).

Figure 3.66. Some figures based on the tear drop.

3.7.9. Pie Charts. A sector is closely related to an arc: each end of the arc is connected to the center of the
circle. The familiar pie chart is formed by drawing a number of sectors. A typical example is shown in Figure
3.67. Pie charts are used to illustrate how a whole is divided into parts, as when a pie is split up and distributed.
The eye quickly grasps how big each slice is relative to the others. Often one or more of the slices is
exploded away from the pack as well, as shown in the figure. Sectors that are exploded are simply shifted
slightly away from the center of the pie chart in the proper direction

Figure 3.67. A pie chart.

To draw a pie chart we must know the relative sizes of the slices. Write and test a routine that accepts data from
the user and draws the corresponding pie chart. The user enters the fraction of the pie each slice represents,
along with an e if the slice is to be drawn exploded, or an n otherwise.

3.8. Using the Parametric form for a curve.

There are two principal ways to describe the shape of a curved line: implicitly and parametrically. The
implicit form describes a curve by a function F(x, y) that provides a relationship between the x and y
coordinates: the point (x, y) lies on the curve if and only if it satisfies:
F(x, y) = 0

condition for (x, y) to lie on the curve

Computer Graphics

Chap 3

09/21/99

(3.7)

5:38 PM

page 37

For example, the straight line through points A and B has implicit form:
F(x, y) = (y - Ay)(Bx - Ax) - (x - Ax)(By - Ay)

(3.8)

and the circle with radius R centered at the origin has implicit form:
2
2
2
F(x, y) = x + y - R

(3.9)

A benefit of using the implicit form is that you can easily test whether a given point lies on the curve:
simply evaluate F(x, y) at the point in question. For certain classes of curves it is meaningful to speak of
an inside and an outside of the curve, in which case F(x, y) is also called the inside-outside function,
with the understanding that
F(x, y) = 0
F(x, y) > 0
F(x, y) < 0

for all (x, y) on the curve

for all (x, y) outside the curve
for all (x, y) inside the curve

(3.10)

(Is F(x, y) of Equation 3.9 a legitimate inside-outside function for the circle?)
Some curves are single-valued in x, in which case there is a function g(.) such that all points on the curve
satisfy y = g(x). For such curves the implicit form may be written F(x, y) = y - g(x). (What is g(.) for the
line of Equation 3.8?) Other curves are single-valued in y, (so there is a function h(.) such that points on
the curve satisfy x = h(y). And some curves are not single-valued at all: F(x, y) = 0 cannot be rearranged
into either of the forms y = g(x) nor x = h(y). The circle, for instance, can be expressed as:

y = R2 x 2

(3.11)

but here there are two functions, not one.

3.8.1. Parametric Forms for Curves.

A parametric form for a curve produces different points on the curve based on the value of a parameter.
Parametric forms can be developed for a wide variety of curves, and they have much to recommend
them, particularly when one wants to draw or analyze the curve. A parametric form suggests the
movement of a point through time, which we can translate into the motion of a pen as it sweeps out the
curve. The path of the particle traveling along the curve is fixed by two functions, x( ) and y( ), and we
speak of (x(t), y(t)) as the position of the particle at time t. The curve itself is the totality of points
visited by the particle as t varies over some interval. For any curve, therefore, if we can dream up
suitable functions x( ) and y( ) they will represent the curve concisely and precisely.
The familiar Etch-a-Sketch8 shown in Figure 3.68 provides a vivid analogy. As knobs are turned, a stylus
hidden in the box scrapes a thin visible line across the screen. One knob controls the horizontal position,
and the other directs the vertical position of the stylus. If the knobs are turned in accordance with x(t) and
y(t), the parametric curve is swept out. (Complex curves require substantial manual dexterity.)
1`st Ed. Figure 7.11
Figure 3.68. Etch-a-Sketch drawings of parametric curves. (Drawing by Suzanne Casiello.)
Examples: The line and the ellipse.
The straight line of Equation 3.8 passes through points A and B. We choose a parametric form that visits
A at t = 0 and B at t = 1, obtaining:
x(t) = Ax + (Bx - Ax)t
y(t) = Ay + (By - Ay)t
8Etch-a-Sketch

(3.12)

is a trademark of Ohio Art.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 38

Thus the point P(t) = (x(t), y(t)) sweeps through all of the points on the line between A and B as t varies
from 0 to 1 (check this out).
Another classic example is the ellipse, a slight generalization of the circle. It is described parametrically
by
x(t) = W cos(t)
y(t) = H sin(t)

(3.13)

, for 0 t 2.

Here W is the half-width, and H the half-height of the ellipse. Some of the geometric properties of
the ellipse are explored in the exercises. When W and H are equal the ellipse is a circle of radius W.
Figure 3.69 shows this ellipse, along with the component functions x(.) and y(.).
y(t)
y
@t = /2
(x(t), y(t))
H
H
@t =
t
W
2
t

-c

x
-H

@t = 3/2
W

x(t)

2
t
Figure 3.69. An ellipse described parametrically.
As t varies from 0 to 2 the point P(t) = (x(t), y(t)) moves once around the ellipse starting (and finishing)
at (W, 0). The figure shows where the point is located at various times t. It is useful to visualize
drawing the ellipse on an Etch-a-Sketch. The knobs are turned back and forth in an undulating pattern,
one mimicking W cos(t) and the other H sin(t). (This is surprisingly difficult to do manually.)
Finding an implicit form from a parametric form - implicitization.
Suppose we want to check that the parametric form in Equation 3.13 truly represents an ellipse. How
do we find the implicit form from the parametric form? The basic step is to combine the two
equations for x(t) and y(t) to somehow eliminate the variable t. This provides a relationship that must
hold for all t. It isnt always easy to see how to do this there are no simple guidelines that apply
for all parametric forms. For the ellipse, however, square both x/W and y/H and use the well-known
fact cos(t)2 + sin(t)2 = 1 to obtain the familiar equation for an ellipse:

x y
W + H =1
2

(3.14)

The following exercises explore properties of the ellipse and other classical curves. They develop
useful facts about the conic sections, which will be used later. Read them over, even if you dont stop to
solve each one.
Practice Exercises

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 39

3.8.1. On the geometry of the Ellipse. An ellipse is the set of all points for which the sum of the
distances to two foci is constant. The point (c, 0) shown in Figure 3.69 forms one focus, and (-c, 0)
forms the other. Show that H, W, and c are related by: W2 = H2 + c2.
3.8.2. How eccentric. The eccentricity, e = c / W, of an ellipse is a measure of how non circular the
ellipse is, being 0 for a true circle. As interesting examples, the planets in our solar system have very
nearly circular orbits, with e ranging from 1/143 (Venus) to 1/4 (Pluto). Earths orbit exhibits e = 1/60.
As the eccentricity of an ellipse approaches 1, the ellipse flattens into a straight line. But e has to get very
close to 1 before this happens. What is the ratio H / W of height to width for an ellipse that has e = 0.99?
3.8..3. The other Conic Sections.
The ellipse is one of the three conic sections, which are curves formed by cutting (sectioning) a circular
cone with a plane, as shown in Figure 3.70. The conic sections are:
ellipse: if the plane cuts one nappe of the cone;
hyperbola: if the plane cuts both nappes
parabola: if the plane is parallel to the side of the cone;

Figure 3.70. The classical conic sections.

The parabola and hyperbola have interesting and useful geometric properties. Both of them have simple
implicit and parametric representations.
Show that the following parametric representations are consistent with the implicit forms given:
Parabola:
Implicit form: y2 - 4 a x = 0
(3.15)
x(t) = a t2
y(t) = 2 a t
Hyperbola:
Implicit form: (x/a)2 - (y/b)2 = 1
(3.16)
x(t)=a sec(t)
y(t)=b tan(t)
What range in the parameter t is used to sweep out this hyperbola? Note: A hyperbola is defined as the
locus of all points for which the difference in its distances from two fixed foci is a constant. If the foci
here are at (-c, 0) and (+c, 0), show that a and b must be related by c2 = a2 + b2.

3.8.2. Drawing curves represented parametrically.

It is straightforward to draw a curve when its parametric representation is available. This is a major
advantage of the parametric form over the implicit form. Suppose a curve C has the parametric
representation P(t) = (x(t), y(t)) as t varies from 0 to T (see Figure 3.71a). We want to draw a good
approximation to it, using only straight lines. Just take samples of P(t) at closely spaced instants. A
sequence {ti} of times are chosen, and for each ti the position Pi = P(ti) = (x(ti), y(ti)) of the curve is
found. The curve P(t) is then approximated by the polyline based on this sequence of points Pi, as shown
in Figure 3.71b.

@t=T

a).

b).
P2

P(t) = (x(t), y(t))

@t=0

Figure 3.71. Approximating a curve by a polyline.

Figure 3.72 shows a code fragment that draws the curve (x(t), y(t)) when the desired array of sample
times t[i] is available.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 40

// draw the curve (x(t), y(t)) using

// the array t[0],..,t[n-1] of sample-times
glBegin(GL_LINES);
for(int i = 0; i < n; i++)
glVertex2f((x(t[i]), y(t[i]));
glEnd();
Figure 3.72. Drawing the ellipse using points equispaced in t.
If the samples are spaced sufficiently close together, the eye will naturally blend together the line
segments and will see a smooth curve. Samples must be closely spaced in t-intervals where the curve is
wiggling rapidly, but may be placed less densely where the curve is undulating slowly. The required
closeness or quality of the approximation depends on the situation.
Code can often be simplified if it is needed only for a specific curve. The ellipse in Equation 3.13 can be
drawn using n equispaced values of t with:
#define TWOPI 2 * 3.14159265
glBegin(GL_LINES);
for(double t = 0; t <= TWOPI; t += TWOPI/n)
glVertex2f(W * cos(t), H * sin(t));
glEnd();
For drawing purposes, parametric forms circumvent all of the difficulties of implicit and explicit forms.
Curves can be multi-valued, and they can self-intersect any number of times. Verticality presents no
special problem: x(t) simply becomes constant over some interval in t. Later we see that drawing curves
that lie in 3D space is just as straightforward: three functions of t are used, and the point at t on the curve
is (x(t), y(t), z(t)).
Practice Exercises.
3.8.4. An Example Curve. Compute and plot by hand the points that would be drawn by the fragment
above for W = 2, H = 1, at the 5 values of t = 2i/9, for i = 0, 1, ...,4.
3.8.5. Drawing a logo. A well-known logo consists of concentric circles and ellipses, as shown in Figure
3.73. Suppose you have a drawing tool: drawEllipse(W, H, color) that draws the ellipse of
Equation 3.13 filled with color color. (Assume that as each color is drawn it completely obscures any
previously drawn color.) Choose suitable dimensions for the ellipses in the logo and give the sequence of
commands required to draw it.

Figure 3.73. A familiar eye made of circles and ellipses.

Some specific examples of curves used in computer graphics will help to cement the ideas.

3.8.3. Superellipses
An excellent variation of the ellipse is the superellipse, a family of ellipse-like shapes that can produce
good effects in many drawing situations. The implicit formula for the superellipse is

x
W

y
+
H

Computer Graphics

(3.17)

Chap 3

09/21/99

5:38 PM

page 41

where n is a parameter called the bulge. Looking at the corresponding formula for the ellipse in Equation
3.14, the superellipse is seen to become an ellipse when n = 2. The superellipse has the following
parametric representation:
x(t) = W cos(t)|cos(t)

2/n -1

|
(3.18)

y(t) = H sin(t)|sin(t)

2/n-1

for 0 t 2. The exponent on the sin() and cos() is really 2/n, but the peculiar form as shown is used
to avoid trying to raise a negative number to a fractional power. A more precise version avoids this.
Check that this form reduces nicely to the equation for the ellipse when n = 2. Also check that the
parametric form for the superellipse is consistent with the implicit equation.
Figure 3.74a shows a family of supercircles, special cases of superellipses for which W = H. Figure
3.74b shows a scene composed entirely of superellipses, suggesting the range of shapes possible.
1st Ed. Figures 4.16 and 4.17 together
Figure 3.74. Family of supercircles. b). Scene composed of superellipses.
For n > 1 the bulge is outward, whereas for n < 1 it is inward. When n = 1, it becomes a square. (In
Chapter 6 we shall look at three-dimensional superquadrics, surfaces that are sometimes used in CAD
systems to model solid objects.)
Superellipses were first studied in 1818 by the French physicist Gabriel Lam. More recently in 1959, the
extraordinary inventor Piet Hein (best known as the originator of the Soma cube and the game Hex) was
approached with the problem of designing a traffic circle in Stockholm. It had to fit inside a rectangle
(with W/ H = 6 / 5) determined by other roads, and had to permit smooth traffic flow as well as be
pleasing to the eye. An ellipse proved to be too pointed at the ends for the best traffic patterns, and so
Piet Hein sought a fatter curve with straighter sides and dreamed up the superellipse. He chose n = 2.5 as
the most pleasing bulge. Stockholm quickly accepted the superellipse motif for its new center. The curves
were strangely satisfying, neither too rounded nor too orthogonal, a happy blend of elliptical and
rectangular beauty [Gardner75, p. 243]. Since that time, superellipse shapes have appeared in furniture,
textile patterns, and even silverware. More can be found out about them in the references, especially in
[Gardner75] and [Hill 79b].
The superhyperbola can also be defined [Barr81]. Just replace cos(t) by sec(t), and sin(t) by tan(y), in
Equation 3.18. When n = 2, the familiar hyperbola is obtained. Figure 3.75 shows example
superhyperbolas. As the bulge n increases beyond 2, the curve bulges out more and more, and as it
decreases below 2, it bulges out less and less, becoming straight for n = 1 and pinching inward for n < 1.
1st Ed. Figure 9.14.
Figure 3.75. The superhyperbola family.

3.8.4. Polar Coordinate Shapes

Polar coordinates may be used to represent many interesting curves. As shown in Figure 3.76, each point
on the curve is represented by an angle and a radial distance r. If r and are each made a function of t,
then as t varies the curve (r(t), (t)) is swept out. Of course this curve also has the Cartesian
representation (x(t), y(t)) where:

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 42

Figure 3.76. Polar coordinates.

x(t) = r(t) cos((t))
y(t) = r(t) sin((t)).

(3.19)

But a simplification is possible for a large number of appealing curves. In these instances the radius r is
expressed directly as a function of , and the parameter that sweeps out the curve is itself. For each
point (r, ) the corresponding Cartesian point (x, y) is given by
x = f() cos()
y = f() sin()

(3.20)

Curves given in polar coordinates can be generated and drawn as easily as any others: The parameter is ,
which is made to vary over an interval appropriate to the shape. The simplest example is a circle with
radius K: f() = K. The form f() = 2K cos() is another simple curve (which one?). Figure 3.77 shows
some shapes that have simple expressions in polar coordinates:
1st Ed. Figure 4.19
Figure 3.77. Examples of curves with simple polar forms..
Cardioid:
Rose curves:
shown.

f() = K (1 + cos()).
f() = K cos(n ), where n specifies the number of petals in the rose. Two cases are

Archimedian spiral:

f() = K.

In each case, constant K gives the overall size of the curve. Because the cardioid is periodic, it can be
drawn by varying from 0 to 2. The rose curves are periodic when n is an integer, and the Archimedian
spiral keeps growing forever as increases from 0. The shape of this spiral has found wide use as a cam
to convert rotary motion to linear motion (see [Yates46] and [Seggern90].
The conic sections (ellipse, parabola, and hyperbola) all share the following polar form:

f ( ) =

1
1 e cos( )

(3.21)

where e is the eccentricity of the conic section. For e = 1 the shape is a parabola; for 0 e <1 it is an
ellipse; and for e > 1 it is a hyperbola.
The Logarithmic Spiral
The logarithmic spiral (or equiangular spiral) f() = Kea, shown in Figure 3.78a, is also of particular
interest [Coxeter61]. This curve cuts all radial lines at a constant angle , where a = cot(). This is the
only spiral that has the same shape for any change of scale: Enlarge a photo of such a spiral any amount,
and the enlarged

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 43

1st Ed Figures 4.20 and 4.21

Figure 3.78. The logarithmic spiral and b). chambered nautilus
spiral will fit (after a rotation) exactly on top of the original. Similarly, rotate a picture of an equiangular
spiral, and it will seem to grow larger or smaller [Steinhaus69]9. This preservation of shape seems to be
used by some animals such as the mollusk inside a chambered nautilus (see Figure 3.78b). As the animal
grows, its shell also grows along a logarithmic spiral in order to provide a home of constant shape
[Gardner61].
Other families of curves are discussed in the exercises and Case Studies, and an exhaustive listing and
characterization of interesting curves is given in [yates46, seggern90, shikin95].

3.8.5. 3D Curves.
Curves that meander through 3D space may also be represented parametrically, and will be discussed
fully in later chapters. To create a parametric form for a 3D curve we invent three functions x(.), y(.), and
z(.), and say the curve is at P(t) = (x(t), y(t), z(t)) at time t.
Some examples are:
The helix: The circular helix is given parametrically by:
x(t) = cos(t)
y(t)= sin(t)
z(t) = bt

(3.22)

for some constant b. It illustrated in Figure 3.79 as a stereo pair. See the Preface for viewing stereo pairs.
If you find this unwieldy, just focus on one of the figures.

Figure 3.79. The helix, displayed as a stereo pair.

Many variations on the circular helix are possible, such as the elliptical helix P(t) = (W cos(t), H sin(t),
bt), and the conical helix P(t) = (t cos(t), t sin(t), bt) (sketch these). Any 2D curve (x(t), y(t)) can of
course be converted to a helix by appending z(t) = bt, or some other form for z(t).
The toroidal spiral. A toroidal spiral, given by
x(t) = (a sin(ct) + b) cos(t)
y(t) = (a sin(ct) + b) sin(t)
z(t) = a cos(ct)

(3.23)

9This

curve was first described by Descartes in 1638. Jacob Bernoulli (1654---1705)

Chap 3

09/21/99

5:38 PM

page 48

private:
Point2 CP;
// current position in the world
IntRect viewport; // the current window
RealRect window; // the current viewport
float mapA, mapB, mapC, mapD; // data for the window to viewport mapping
void makeMap(void); // builds the map
int screenWidth, screenHeight;
float delx,dely;
// increments for clipper
char code1, code2;
// outside codes for clipper
void ChopLine(tPoint2 &p, char c);
int clipSegment(tPoint2 &p1, tPoint2 &p2);
};
Figure 3.86. Interface for the Canvas class for Turbo C++.

Implementation of the Canvas class.

We show some of the Canvas member functions here, to illustrate what must be done to manage the window to
viewport mapping and clipping ourselves.
1). The Canvas constructor.
The constructor is passed the desired width and height of the screen. Turbo C++ is placed in graphics mode at
the highest resolution supported by the graphics system. The actual screen width and height available is tested,
and if it is less than was requested, the program terminates. Then a default window and viewport are established,
and the window to viewport mapping is built (inside setViewport().)
Canvas:: Canvas(int width, int height)
{
int gdriver = DETECT, gmode; //Turbo C++ : use best resolution screen
initgraph(&gdriver, &gmode, ""); // go to graphics mode
screenWidth = getmaxx() + 1; // size of available screen
screenHeight = getmaxy() + 1;
assert(screenWidth >= width); // as wide as asked for?
assert(screenHeight >= height); // as high as asked for?
CP.set(0.0, 0.0);
window.set(-1.0,1.0,-1.0,1.0); // default window
setViewport(0,screenWidth, 0, screenHeight); // sets default map too
}
2). Setting the window and viewport and the mapping.
Whenever either the window or viewport is set, the window to viewport mapping is updated to insure that it is current.
A degenerate window of zero height causes an error. The mapping uses window and viewport data to compute the four
coefficients A, B, C, and D required.
//<<<<<<<<<<<<<<<<<<<<<<<< set Window >>>>>>>>>>>>>>>>>>>>
void Canvas:: setWindow(float l, float r, float b, float t)
{
window.set(l, r, b, t);
assert(t != b); //degenerate !
makeMap();
// update the mapping
}
//<<<<<<<<<<<<<<<<<<<< setViewport >>>>>>>>>>>>>>>>>>>
void Canvas:: setViewport(int l, int r, int b, int t)
{
viewport.set(l, r, b, t);
makeMap();
// update the mapping
}
//<<<<<<<<<<<<<<<< makeMap >>>>>>>>>>>>>>>>>>>>>
void Canvas:: makeMap(void)
{
// set mapping from window to viewport

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 49

intRect vp = getViewport(); // local copy of viewport

RealRect win = getWindow(); // local copy of window
float winWid = win.r - win.l;
float winHt = win.t - win.b;
assert(winWid != 0.0); assert(winHt != 0.0); // degenerate!
mapA = (vp.r - vp.l)/winWid; // fill in mapping values
mapB = vp.l - map.A * win.l;
mapC = (vp.t - vp.b)/winHt;
mapD = vp.b - map.B * win.b;
}
3). moveTo(), and lineTo() with clipping.
The routine moveTo() converts its point from world coordinates to screen coordinates, and calls the Turbo C++
specific moveto() to update the internal current position maintained by Turbo C++. It also updates Canvas
world coordinate CP. Routine lineTo() works similarly, but it must first determine which part if any of the
segment lies within the window. To do this it uses clipSegment() described in Section 3.3 and in Case Study
3.2, which returns the first and second endpoints of the inside portion. If so it moves to first and draws a
line to second. It finishes with a moveTo() to insure that the CP will be current (both the Canvas CP and the
internal Turbo C++ CP).
ChopLine and ClipSegment are same as in Case Study 3.2.
//<<<<<<<<<<<<<<<<<<<<<<<<<< moveTo >>>>>>>>>>>>>>>>>>
void Canvas:: moveTo(float x, float y)
{
int sx = (int)(mapA * x + mapC);
int sy = (int)(mapB * y + mapD);
moveto(sx, sy);
// a Turbo C++ routine
CP.set(x, y);
}
//<<<<<<<<<<<<<<<<<<<< lineTo >>>>>>>>>>>
void Canvas:: lineTo(float x, float y)
{ // Draw a line from CP to (x,y), clipped to the window
Point2 first = CP; // initial value of first
Point2 second(x, y); // initial value of second
if(clipSegment(first, second)) // any part inside?
{
moveTo(first.x, first.y); // to world CP
int sx = (int)(mapA * second.x + mapC);
int sy = (int)(mapB * second.y + mapD);
lineto(sx,sy); // a Turbo C++ routine
}
moveTo(x, y);
// update CP
}
Write a full implementation of the Canvas class for Turbo C++ (or a similar environment that requires you to
implement clipping and mapping). Cope appropriately with setting the drawing and background colors (this is
usually quite system-specific). Test your class by using it in an application that draws polyspirals as specified by
the user.

3.10.4. Case Study 3.4. Drawing Arches.

(Level of Effort: II) Arches have been used throughout history in architectural compositions. Their structural
strength and ornamental beauty make them very important elements in structural design, and a rich variety of
shapes have been incorporated into cathedrals, bridges, doorways, etc.
Figure 3.87 shows two basic arch shapes. The arch in part a) is centered at the origin, and has a width of 2W .
The arch begins at height H above the base line. Its principal element is a half-circle with a radius R = W. The
ratio H/W can be adjusted according to taste. For instance, H/W might be related to the golden ratio.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 50

a) Rounded Arch

b). Pointed Arch

R
R
W

-H
-H

Figure 3.87. Two basic arch forms.

Figure 3.73b shows an idealized version of the second most famous arch shape, the pointed or equilateral
arch, often seen in cathedrals10. Here two arcs of radius R = 2W meet directly above the center. (Through what
angle does each arc sweep?)
The ogee11 (or keel) arch is shown in Figure 3.88. This arch was introduced about 1300 AD, and was popular
in architectural structures throughout the late Middle Ages. Circles of radius f R rest on top of a rounded arch of
radius R for some fraction f. This fixes the position of the two circles. (What are the coordinates of point C?)
On each side two arcs blend together to form a smooth pointed top. It is interesting to work out the parameters
of the various arcs in terms of W and f.
y

C
fR

R
W

Figure 3.88. The Ogee arch.

Develop routines that can draw each of the arch types described above. Also write an application that draws an
interesting collection of such arches in a castle, mosque, or bridge of your design.

3.10.5. Case Study 3.5. Some Figures used in Physics and Engineering.
(Level of Effort: II) This Case Study works with a collection of interesting pictures that arise in certain topics
within physics and engineering. The first illustrates a physical principal of circles intersecting at right angles;
the second creates a chart that can be used to study electromagnetic phenomena; the third develops symbols that
are used in designing digital systems.
1). Electrostatic Fields. The pattern of circles shown in Figure 3.89 is studied in physics and electrical
engineering, as the electrostatic field lines that surround electrically charged wires. It also appears in
mathematics in connection with the analytic functions of a complex variable. In Chapter 5 these families also
are found when we examine a fascinating set of transformations, inversions in a circle. Here we view them
simply as an elegant array of circles and consider how to draw them.
10From J.Fleming, H. Honour, N. Pevsner: Dictionary of Architecture. Penguin Books, London 1980
11From the old French ogive meaning an S-shaped curve.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 51

two-pointers
surrounders

-a

Figure 3.89. Families of orthogonal circles..

There are two families of circles, which we will call two-pointers and surrounders. The two-pointers family
consists of circles that pass through two given points. Suppose the two points are (-a, 0) and (a, 0). The twopointers can be distinguished by some parameter m, and for each value of m two different circles are generated
(see Figure 3.75). The circles have centers and radii given by:
center= (0, a

m2 - 1 ) and

radius = am

as m varies from 1 to infinity.

Circles in the surrounders family surround one of the points (-a, 0 ) or ( a, 0). The centers and radii of the
surrounders are also distinguished by a parameter n and have the values
center = ( an, 0) and radius = a n2-1
as n varies from 1 to infinity. The surrounder circles are also known as circles of Appolonius, and they arise
in problems of pursuit [Ball & Coxeter]. The distances from any point on a circle of Appolonius to the points
(-a, 0) and (a, 0) have a constant ratio. (What is this ratio in terms of a and n?)
The surrounder family is intimately related to the two-pointer family: Every surrounder circle cuts through
every two-pointer circle at a right angle. The families of circles are thus said to be orthogonal to one another.
Write and exercise a program that draws the two families of orthogonal circles. Choose sets of values of m and
n so that the picture is well balanced and pleasing.
2). Smith Charts. Another pattern of circles is found in Smith charts, familiar in electrical engineering in
connection with electromagnetic transmission lines. Figure 3.90 shows the two orthogonal families found in
Smith charts. Here all members of the families pass through a common point (1, 0). Circles in family A have
centers at (1 -m, 0) and radii m, and circles in family B have centers at (1, n) and radii n, where both m and n
vary from 0 to . Write and exercise a program that draws these families of circles.
1st Ed. Figure 4.31
Figure 3.90. The Smith Chart.
3). Logic Gates for Digital Circuits. Logic gates are familiar to scientists and engineers who study basic
electronic circuits found in computers. Each type of gate is symbolized in a circuit diagram by a characteristic
shape, several of which are based on arcs of circles. Figure 3.91a shows the shape of the so-called

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 52

a). NAND gate

b). NOR gate

26
26

13
32

Figure 3.91. Standard Graphic Symbol for the Nand and Nor Gates.
NAND gate, according to a world-wide standard12. The NAND gate is basically a rounded arch placed on its
side. The arc has radius 13 units relative to the other elements, so the NAND gate must be 26 units in height.
Figure 3.91b shows the standard symbol for a NOR gate. It is similar to a pointed arch turned on its side. Three
arcs are used, each having a radius of 26 units. (The published standard as shown has an error in it, that makes it
impossible for certain elements to fit together. What is the error?)
Write a program that can draw both of these circuit types at any size and position in the world. (For the NOR
gate find and implement a reasonable correction to the error in Figure 3.77b.) Also arrange matters so that your
program can draw these gates rotated by 90 , 180, or 270.

3.10.6. Case Study 3.6. Tilings.

(Level of Effort: II) Computer graphics offers a powerful tool for creating pleasing pictures based on geometric
objects. One of the most intriguing types of pictures are those that apparently repeat forever in all directions.
They are called variously tilings, and repeat patterns. They are studied in greater detail in Chapter ???.
A). Basic Tilings. Figure 3.92 shows a basic tiling. A motif, in this case four quarter circles in a simple
arrangement, is designed in a square region of the world. To draw a tiling over the plane based on this motif, a
collection of viewports are created side by side that cover the display surface, and the motif is drawn once
inside each viewport.
a).

b).

Figure 3.92. A motif and the resulting tiling.

Write a program that:
a). chooses a square window in the world, and draws some interesting motif (possibly clipping portions of it, as
in Figure 3.14);
b). successively draws the picture in a set of viewports that abut one another and together cover the display
surface.
Exercise your program with at least two motifs.

12The Institute of Electrical and Electronic Engineers (IEEE) publishes many things, including standard definitions

of terminology and graphic shapes of circuit elements. These drawings are taken from the standard document: IEEE
Std. 91-1984 .

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 53

B). Truchet Tiles. A slight variation of the method above selects successive motifs randomly from a pool of
candidate motifs. Figure 3.93a shows the well-known Truchet tiles13, which are based on two quarter circles
centered at opposite corners of a square. Tile 0 and tile 1 differ only by a 90 rotation.

artist sketches the two tiles here

Figure 3.93. Truchet Tiles. a). the two tiles. b). A truchet pattern

.
Figure 3.93.b.
Write an application that draws Truchet tiles over the entire viewport. Each successive tile uses tile 0 or tile 1,
selected at random.
Curves other than arcs can be used as well, as suggested in Figure 3.94. What conditions should be placed on
the angle with which each curve meets the edge of the tile in order to avoid sharp corners in the resulting curve?
This notion can also be extended to include more than two tiles.

Figure 3.94. Extension of Truchet tiles.

Extend the program above so that it introduces random selections of two or more motifs, and exercise it on the
motifs you have designed. Design motifs that blend together properly.

3.10.7. Case Study 3.7. Playful Variations on a Theme.

(Estimate of time required: four hours). In Section 3.8 we discussed how to draw a curve represented
parametrically by P(t): take a succession of instants {ti} and connect the successive "samples" (x(ti), y(ti))
by straight lines. A wide range of pictures can be created by varying the way in which the samples are
taken. We suggest some possibilities here.
13Smith, C. The Tiling Patterns of Sebastian Truchet and the topology of structural hierarchy. Leonardo, 20:4, pp:

373-385, 1987. (refd in Pickover, p.386)

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 54

Write a program that draws each of the four shapes:

a). an ellipse
b). a hyperbola
c). a logarithmic spiral
d). a 5-petal rose curve
for each of the methods described below for obtaining t-samples.
1). Unevenly Spaced Values of t. Instead of using a constant increment between values of t when
sampling the functions x() and y(), use a varying increment. It is interesting to experiment with different
choices to see what visual effects can be achieved. Some possibilities for a sequence of (n+1) t-values
between 0 and T (suitably chosen for the curve shape at hand) are
ti = T i / n : The samples cluster closer and closer together as i increases.
ti = T (i / n) : The samples spread out as i increases.
ti = T (i / n) + Asin(ki / n) The samples cyclically cluster together or spread apart. Constants A and
k are chosen to vary the amount and speed of the variation.
2

2). Randomly Selected t-Values. The t-values can be chosen randomly as in

ti = randChoose(0,T)
Here randChoose(0,T) is a function (devised by you) that returns a value randomly selected from the
range 0 to T each time it is called. (See Appendix 3 for a basic random number generator.)
Figure 3.95 shows the polyline generated in this fashion for points on an ellipse. It is interesting to watch
such a picture develop on a display. A flurry of seemingly unrelated lines first appears, but soon the eye
detects some order in the chaos and sees an elliptical envelope emerging around the cloud of lines.
1st Ed. Figure 4.26
Figure 3.95. A random ellipse polyline.
Alternatively, a sequence of increasing t-values can be used, generated by
ti = ti-1 + randChoose(0, r)
where r is some small positive value.
3). Connecting Vertices in Different Orders
In a popular childrens game, pins are driven into a board in some pattern, and a piece of thread is woven
around the pins in some order. The t-values here define the positions of the pins in the board, and
worldLineTo() plays the role of the thread.
The samples of P(t) are prestored in a suitable array P[i], i = 0,1,..,n.The polyline is drawn by
sequencing in an interesting way through values of i. That is, the sequence i0, i1, .... is generated from
values between 0 and n, and for each index ik a call to worldLineTo(P[ik]) is made. Some
possibilities are:
Random Deal: the sequence i0, i1, .... is a random permutation of the values 0,1,..,n, as in dealing a
fixed set of cards from a shuffled deck.
Every pair of points is connected by a straight line. So every pair of values in the range 0,1,..,n appears
in adjacent spots somewhere in the sequence i0, i1, .... The prime rosette of Chapter 5 gave one example,
where lines were drawn connecting each point to every other.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 55

One can also draw webs, as suggested in Figure 3.96. Here the index values cycle many times
through the possible values, skipping by some M each time. This is easily done by forming the next index
from the previous one using i = (i + M) mod (n+1).
1st Ed. Figure 4.27
Figure 3.96. Adding webs to a curve.

3.10.8. Case Study 3.8. Circles Rolling around Circles.

(Level of Effort: II) Another large family of interesting curves can be useful in graphics. Consider the
path traced by a point rigidly attached to a circle as the circle rolls around another fixed circle [thomas53,
Yates46]. These are called trochoids, and Figure 3.97 shows how they are generated. The tracing point is
attached to the rolling circle (of radius b) at the end of a rod k units from the center. The fixed circle has
radius a. There are two basic kinds: When the circle rolls externally (Figure 3.97a), an epitrochoid is
generated, and when it rolls internally (Figure 3.97b), a hypotrochoid is generated. The childrens game
Spirograph14 is a familiar tool for drawing trochoids, which have the following parametric forms:
1st Ed. Figure 4.23
Figure 3.97. Circles rolling around circles.
The epitrochoid:

(a + b)t
)
b
( a + b)t
)
y(t ) = (a + b)sin(2 t ) k sin(2
b
x (t ) = ( a + b )cos(2 t ) k cos(2

(3.24)

The hypotrochoid:

(a b)t
)
b
( a b )t
)
y(t ) = (a b)sin(2 t ) k sin(2
b
x (t ) = ( a b )cos(2 t ) + k cos(2

(3.25)

An ellipse results from the hypotrochoid when a = 2b for any k.

When the tracing point lies on the rolling circle ( k = b) these shapes are called cycloids. Some familiar
special cases of cycloids are
Epicycloids:
Cardioid: b = a
Nephroid: 2b = a
Hypocycloids:15
Line segment: 2b = a
Deltoid: 3b = a
Astroid: 4b = a.
Some of these are shown in Figure 3.98. Write a program that can draw both epitrochoids and
hypotrochoids. The user can choose which family to draw, and can enter the required parameters.
Exercise the program to draw each of the special cases listed above.
1st Ed. Figure 4.24.
Figure 3.98. Examples of cycloids: a) nephroid, b) a/b = 10, c) deltoid, d) astroid.

14
15

A trademark of Kenner Products.

Note that the astroid is also a superellipse! It has a bulge of 2/3.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 56

3.10.9. Case Study 3.9. Superellipses.

(Level of Effort: I) Write and exercise a program to draw superellipses. To draw each superellipse, the
user indicates opposite corners of its bounding, and types a value for the bulge, whereupon the specified
superellipse is drawn.
(Optional). Extend the program so that it can draw rotated superellipses. The user types an angle after
typing the bulge.

3.11. For Further Reading.

When getting started with graphics it is very satisfying to write applications that produce fascinating curves and
patterns. This leads you to explore the deep connection between mathematics and the visual arts. Many books
are available that offer guidance and provide myriad examples. McGregor and Watts THE ART OF
GRAPHICS FOR THE IBM PC, [mcgregor86] offers many algorithms for creating interesting patterns. Some
particularly noteworthy books on curves and geometry are Jay Kappraffs CONNECTIONS [kappraff91],
Dewdneys THE ARMCHAIR UNIVERSE [dewdney88], Stan Ogilvys EXCURSIONS IN
GEOMETRY[ogilvy69], Pedoes GEOMETRY AND THE VISUAL ARTS [pedoe76], Roger Sheperds MIND
SIGHTS [shep90], and the series of books on mathematical excursions by Martin Gardner, (such as TIME
TRAVEL [gardner88] and PENROSE TILES TO TRAPDOOR CIPHERS [gardner89]). Coxeter has written
elegant books on geometry, such as INTRODUCTION TO GEOMETRY[Coxeter69] and MATHEMATICAL
RECREATIONS AND ESSAYS [ball74], and Hoggars MATHEMATICS FOR COMPUTER GRAPHICS
[hoggar92] discusses many features of iterated function systems.

Computer Graphics

Chap 3

09/21/99

5:38 PM

page 57

(for ECE660, Fall, 1999)

Chapter 4. Vectors Tools for Graphics.

The knowledge at which geometry aims is knowledge of the eternal,
and not of aught perishing and transient.
Plato
For us, whose shoulders sag under the weight of the heritage of Greek thought
and who walk in the paths traced out by the heroes of the Renaissance,
a civilization without mathematics is unthinkable.
Andre Weil
Let us grant that the pursuit of mathematics is a divine madness of the human spirit.
Alfred North Whitehead
All that transcend geometry, transcends our comprehension.
Blaise Pascal

Goals of the Chapter

To review vector arithmetic, and to relate vectors to objects of interest in graphics.

To relate geometric concepts to their algebraic representations.
To describe lines and planes parametrically.
To distinguish points and vectors properly.
To exploit the dot product in graphics topics.
To develop tools for working with objects in 3D space, including the cross product of two vectors.

Preview
This chapter develops a number of useful tools for dealing with geometric objects encountered in
computer graphics. Section 4.1 motivates the use of vectors in graphics, and describes the principal
coordinate systems used. Section 4.2 reviews the basic ideas of vectors, and describes the key operations
that vectors allow. Although most results apply to any number of dimensions, vectors in 2D and 3D are
stressed. Section 4.3 reviews the powerful dot product operation, and applies it to a number of geometric
tasks, such as performing orthogonal projections, finding the distance from a point to a line, and finding
the direction of a ray reflected from a shiny surface. Section 4.4 reviews the cross product of two
vectors, and discusses its important applications in 3D graphics.
Section 4.5 introduces the notion of a coordinate frame and homogeneous coordinates, and stresses that
points and vectors are significantly different types of geometric objects. It also develops the two principal
mathematical representations of a line and a plane, and shows where each is useful. It also introduces
affine combinations of points and describes an interesting kind of animation known as tweening. A
preview of Bezier curves is described as an application of tweening.
Section 4.6 examines the central problem of finding where two line segments intersect, which is vastly
simplified by using vectors. It also discusses the problem of finding the unique circle determined by three
points. Section 4.7 discusses the problem of finding where a ray hits a line or plane, and applies the
notions to the clipping problem. Section 4.8 focuses on clipping lines against convex polygons and
polyhedra, and develops the powerful Cyrus-Beck clipping algorithm.
The chapter ends with Case Studies that extend these tools and provide opportunities to enrich your
graphics programming skills. Tasks include processing polygons, performing experiments in 2D ray
tracing, drawing rounded corners on figures, animation by tweening, and developing advanced clipping
tools.

4.1 Introduction.

Hill - Chapter 4

09/23/99

page 1

In computer graphics we work, of course, with objects defined in a three dimensional world (with 2D objects
and worlds being just special cases). All objects to be drawn, and the cameras used to draw them, have shape,
position, and orientation. We must write computer programs that somehow describe these objects, and describe
how light bounces around illuminating them, so that the final pixel values on the display can be computed.
Think of an animation where a camera flies through a hilly scene containing various buildings, trees, roads, and
cars. What does the camera see? It all has to be converted ultimately to numbers. Its a tall order.
The two fundamental sets of tools that come to our aid in graphics are vector analysis and transformations. By
studying them in detail we develop methods to describe the various geometric objects we will encounter, and
we learn how to convert geometric ideas to numbers. This leads to a collection of crucial algorithms that we can
call upon in graphics programs.
In this chapter we examine the fundamental operations of vector algebra, and see how they are used in graphics;
transformations are addressed in Chapter 5. We start at the beginning and develop a number of important tools
and methods of attack that will appear again and again throughout the book. If you have previously studied
vectors much of this chapter will be familiar, but the numerous applications of vector analysis to geometric
situations should still be scrutinized. The chapter might strike you as a mathematics text. But having it all
collected in one place, and related to the real problems we encounter in graphics, may be found useful.
Why are vectors so important?
A preview of some of some situations where vector analysis comes to the rescue might help to motivate the
study of vectors. Figure 4.1 shows three geometric problems that arise in graphics. Many other examples could
be given.
a ).

b ).

c).

(4,6 )
ce n te r?

(2,2 )

(7,1 )

vie w pla n e

Figure 4.1. Three sample geometric problems that yield readily to vector analysis.
Part a) shows a computer-aided design problem: the user has placed three points on the display with the mouse,
and wants to draw the unique circle that passes through them. (Can you visualize this circle?). For the
coordinates given where is the center of the circle located? We see in Section 4.6 that this problem is thorny
without the use of vectors, but almost trivial when the right vector tools are used.
Part b) shows a camera situated in a scene that contains a Christmas tree. The camera must form an image of the
tree on its viewplane (similar to the film plane of a physical camera), which will be transferred to a screen
window on the users display. Where does the image of the tree appear on this plane, and what is its exact
shape? To answer this we need a detailed study of perspective projections, which will be greatly aided by the
use of vector tools. (If this seems too easy, imagine that you are developing an animation, and the camera is
zooming in on the sphere along some trajectory, and rotating as it does so. Write a routine that generates the
whole sequence of images!)
Part c) shows a shiny cone in which the reflection of a cube can be seen. Given the positions of the cone, cube,
and viewing camera, where exactly does the reflected image appear, and what is its color and shape? When
studying ray tracing in Chapter 15 we will make extensive use of vectors, and we will see that this problem is
readily solved.
Some Basics.
All points and vectors we work with are defined relative to some coordinate system. Figure 4.2 shows the
coordinate systems that are normally used. Each system has an origin called and some axes emanating from
. The axes are usually oriented at right angles to one another. Distances are marked along each axis, and a

Hill - Chapter 4

09/23/99

page 2

point is given coordinates according to how far along each axis it lies. Part a) shows the usual two-dimensional
system. Part b) shows a right handed 3D coordinate system, and part c) shows a left handed coordinate system.
y
a).

b).

(2, 3 )

x
x

Figure 4.2. The familiar two- and three-dimensional coordinate systems.

In a right handed system, if you rotate your right hand around the z-axis by sweeping from the positive xaxis around to the positive y-axis, as shown in the figure, your thumb points along the positive z-axis. In a
left handed system, you must do this with your left hand to make your thumb point along the positive zaxis. Right-handed systems are more familiar and are conventionally used in mathematics, physics, and
engineering discussions. In this text we use a right-handed system when setting up models for objects. But
left-handed systems also have a natural place in graphics, when dealing with viewing systems and
cameras.
We first look at the basics of vectors, how one works with them, and how they are useful in graphics. In
Section 4.5 we return to fundamentals and show an important distinction between points and vectors that, if
ignored, can cause great difficulties in graphics programs.

4.2. Review of Vectors.

Not only Newtons laws, but also the other laws of physics, so far as we know today, have the two
properties which we call invariance under translation of axes and rotation of axes. These properties are so
important that a mathematical technique has been developed to take advantage of them in writing and using
physical laws.. called vector analysis.
Richard Feynman
Vector arithmetic provides a unified way to express geometric ideas algebraically. In graphics we work with
vectors of two, three, and four dimensions, but many results need only be stated once and they apply to
vectors of any dimension. This makes it possible to bring together the various cases that arise in graphics
together into a single expression, which can be applied to a broad variety of tasks.
Viewed geometrically, vectors are objects having length and direction. They correspond to various physical
entities such as force, displacement, and velocity. A vector is often drawn as an arrow of a certain length
pointing in a certain direction. It is valuable to think of a vector geometrically as a displacement from one
point to another.
Figure 4.3 uses vectors to show how the stars in the Big Dipper are moving over time [kerr79]. The current
location of each star is shown by a point, and a vector shows the velocity of each star. The tip of each
arrow shows the point where its star will be located in 50,000 years: producing a very different Big Dipper
indeed!

Hill - Chapter 4

09/23/99

page 3

Figure 4.3. The Big Dipper now and in AD 50,000.

Figure 4.4a shows, in a 2D coordinate system, the two points P = (1, 3) and Q = (4, 1). The displacement
from P to Q is a vector v having components (3, -2),1 calculated by subtracting the coordinates of the points
individually. To get from P to Q we shift down by 2 and to the right by 3. Because a vector is a
displacement it has size and direction but no inherent location: the two arrows labeled v in the figure are in
fact the same vector. Figure 4.4b shows the corresponding situation in three dimensions: v is the vector from
point P to point Q. One often states:
b).

a).

4
3

v
P

v
v

Q
1

x
5

Figure 4.4. A vector as a displacement.

The difference between two points is a vector: v = Q - P;
Turning this around, we also say that a point Q is formed by displacing point P by vector v; we say that v
offsets P to form Q. Algebraically, Q is then the sum: Q = P + v.
The sum of a point and a vector is a point: P + v = Q.
At this point we represent a vector through a list of its components: an n-dimensional vector is given by an
n-tuple:
w = (w1, w2, . . . , wn)

(4.1)

Mostly we will be interested in 2D or 3D vectors as in r = (3.4, -7.78) or t = (33, 142.7, 89.1). Later when it
becomes important we will explore the distinction between a vector and its representation, and in fact will
use a slightly expanded notation to represent vectors (and points). Writing a vector as a row matrix like t =
(33, 142.7, 89.1) fits nicely on the page, but when it matters we will instead write vectors as column
matrices:

33
3.4

r=
.7
7.78 , or t = 142

89.1
It matters when we want to multiply a point or a vector by a matrix, as we shall see in Chapter 5.

4.2.1. Operations with vectors.

Vectors permit two fundamental operations: you can add them, and you can multiply them by scalars (real
numbers)2. So if a and b are two vectors, and s is a scalar, it is meaningful to form both a + b and the
product sa. For example, if a = (2, 5, 6) and b = (-2, 7, 1), we can form the two vectors:
a + b = (0, 12, 7)
1Upper case letters are conventionally used for points, and boldface lower case letters for vectors.
2 There are also systems where the scalars can be complex numbers; we do not work with them here.

Hill - Chapter 4

09/23/99

page 4

6 a = (12, 30, 36)

always performing the operations componentwise. Figure 4.5 shows a two-dimensional example, using a =
(1, -1) and b = (2, 1). We can represent the addition of two vectors graphically in two different ways. In
Figure 4.5a we show both vectors starting at the same point, thereby forming two sides of a parallelogram.
The sum of the vectors is then a diagonal of this parallelogram, the diagonal that emanates from the binding
point of the vectors. This view the parallelogram rule for adding vectors is the natural picture for
forces acting at a point: The diagonal gives the resultant force.
a).

b).

a+b
a+b
b

a
a

Figure 4.5. The sum of two vectors.

Alternatively, in Figure 4.5b we show one vector starting at the head of the other (i.e., place the tail of b at
the head of a) and draw the sum as emanating from the tail of a to the head of b. The sum completes the
triangle, which is the simple addition of one displacement to another. The components of the sum are clearly
the sums of the components of its parts, as the algebra dictates.
Figure 4.6 shows the effect of scaling a vector. For s = 2.5 the vector s a has the same direction as a but is
2.5 times as long. When s is negative, the direction of s a is opposite that of a: The case s = -1 is shown in
the figure.
a
2.5 a
-a

Figure 4.6. Scaling a vector.

Subtraction follows easily once adding and scaling have been established: a - c is simply a + (-c). Figure 4.7
shows the geometric interpretation of this operation, forming the difference of a and c as the sum of a and -c
(Figure 4.7b). Using the parallelogram rule, this sum is seen to be equal to the vector that

a).

b).

c).
a-c

a
c

a-c
a

-c

Figure 4.7. Subtracting vectors.

emanates from the head of c and terminates at the head of a (Figure 4.7c). This is recognized as one diagonal
of the parallelogram constructed using a and c. Note too that it is the other diagonal from the one that
represents the sum a + c.

4.2.2. Linear Combinations of Vectors.

With methods in hand for adding and scaling vectors, we can define a linear combination of vectors. To
form a linear combination of two vectors, v and w, (having the same dimension) we scale each of them by
some scalars, say a and b, and add the weighted versions to form the new vector, a v + b w. The more
general definition for combining m such vectors is:

Hill - Chapter 4

09/23/99

page 5

Definition:
A linear combination of the m vectors v1, v2, . . . , vm is a vector of the form
w = a1 v1 + a2 v2 +. . . + am vm
where a1, a2, . . . , am are scalars.

(4.2)

For example, the linear combination 2(3, 4,-1) + 6(-1, 0, 2) forms the vector (0, 8, 10). In later chapters we
shall deal with rather elaborate linear combinations of vectors, especially when representing curves and
surfaces using spline functions.
Two special types of linear combinations, affine and convex combinations, are particularly important in
graphics.

Affine Combinations of Vectors.

A linear combination is an affine combination if the coefficients a1, a2, . . . , am add up to 1. Thus the linear
combination in Equation 4.2 is affine if:
a1 + a2 + ... + am = 1

(4.3)

For example, 3 a + 2 b - 4 c is an affine combination of a, b, and c, but 3 a + b - 4 c is not. The

coefficients of an affine combination of two vectors a and b are often forced to sum to 1 by writing one as
some scalar t and the other as (1-t):
(1-t) a + (t) b

(4.4)

Affine combinations of vectors appear in various contexts, as do affine combinations of points, as we see
later.

Convex Combinations of Vectors.

Convex combinations have an important place in mathematics, and numerous applications in graphics. A
convex combination arises as a further restriction on an affine combination. Not only must the
coefficients of the linear combination sum to one; each one must also be nonnegative. The linear
combination of Equation (4.2.2) is convex if:
a1 + a2 + ... + am = 1,

(4.5)

and ai 0, for i = 1,,m.. As a consequence all ai must lie between 0 and 1. (Why?).
Thus .3a+.7b is a convex combination of a and b, but 1.8a -.8b is not. The set of coefficients a1, a2, . . . , am
is sometimes said to form a partition of unity, suggesting that a unit amount of material is partitioned into
pieces. Convex combinations frequently arise in applications when one is making a unit amount of some
brew and can combine only positive amounts of the various ingredients. They appear in unexpected
contexts. For instance, we shall see in Chapter 8 that spline curves are in fact convex combinations of
certain vectors, and in our discussion of color in Chapter 12 we shall find that colors can be considered as
vectors, and that any color of unit brightness may be considered to be a convex combination of three primary
colors!
We will find it useful to talk about the set of all convex combinations of a collection of vectors. Consider
the set of all convex combinations of the two vectors v1 and v2. It is the set of all vectors
v = (1 - a) v1 + a v2

(4.6)

as the parameter a is allowed to vary from 0 to 1 (why?) What is this set? Rearranging the equation, v is seen
to be:

Hill - Chapter 4

09/23/99

page 6

v = v1 + a (v2 - v1)

(4.7)

Figure 4.8a shows this to be the vector that is v1 plus some fraction of v2 - v1, so the tip of v lies on the line
joining v1 and v2. As a varies from 0 to 1, v takes on all the positions on the line from v1 to v2, and only
those.
a).
b).
v1

L
a (v2 - v1)

v
v1

.2v1

.3v3
.5v2

Figure 4.8. The set of vectors representable by convex combinations.

Figure 4.8b shows the set of all convex combinations of three vectors. Choose two parameters a1 and a2,
both lying between 0 and 1, and form the following linear combination:
q = a1 v1 + a2 v2 + (1 - a1 - a2 ) v3

(4.8)

where we also insist that a1 plus a2 does not exceed one. This is a convex combination, since none of the
coefficients is ever negative and they sum to one. Figure 4.9 shows the three position vectors v1 = (2, 6), v2 =
(3, 3), and v3 = (7, 4). By the proper choices of a1 and a2, any vector lying within the shaded triangle of
vectors can be represented, and no vectors outside this triangle can be reached. The vector b = .2 v1 + .5 v2 +
.3 v3, for instance, is shown explicitly as the vector sum of the three weighted ingredients. (Note how it is
built up out of portions of the three constituent vectors.) So the set of all convex combinations of these
three vectors spans the shaded triangle. The proof of this is requested in the exercises.
If a2 = 0, any vector in the line L that joins v1 and v3 can be reached by the proper choice of a1. For
example, the vector that is 20 percent of the way from v1 to v3 along L is given by .8 v1+ 0 v2 +.2 v3.

4.2.3. The Magnitude of a vector, and unit vectors.

If a vector w is represented by the n-tuple (w1, w2, . . . , wn), how might its magnitude (equivalently, its
length or size) be defined and computed? We denote the magnitude by |w| and define it as the distance from
its tail to its head. Based on the Pythagorean theorem, this becomes

w = w12 + w2 2 +...+ wn 2

(4.9)

For example, the magnitude of w = (4, -2) is 20 , and that of w = (1, -3, 2) is 14 . A vector of zero
length is denoted as 0. Note that if w is the vector from point A to point B, then |w| will be the distance from
A to B (why?).
It is often useful to scale a vector so that the result has a length equal to one. This is called normalizing a
vector, and the result is known as a unit vector. For example, we form the normalized version of a, denoted
a , by scaling it with the value 1/|a|:

Hill - Chapter 4

09/23/99

page 7

a =

a
a

(4.10)

Clearly this is a unit vector: | a | = 1 (why?), having the same direction as a. For example, if a = (3, -4), then

3 -4
| a | = 5 and the normalized version is a = ( 5 , 5 ) . At times we refer to a unit vector as a direction. Note

is the normalized version of a, vector

that any vector can be written as its magnitude times its direction: If a

a may always be written a =| a | a.
Practice Exercises.
4.2.1. Representing Vectors as linear combinations. With reference to Figure 4.9, what values, or range of
values, for a1 and a2 create the following sets?
a. v1.
b. The line joining v1 and v2.
c. The vector midway between v2 and v3.
d. The centroid of the triangle.
4.2.2. The set of all convex combinations. Show that the set of all convex combinations of three vectors v1, v2,
and v3 is the set of vectors whose tips lie in the triangle formed by the tips of the three vectors. Hint: Each
point in the triangle is a combination of v1 and some point lying between v2 and v3.
4.2.3. Factoring out a scalar. Show how scaling a vector v by a scalar s changes its length. That is, show that:
| s v | = | s | | v | . Note the dual use of the magnitude symbol | |, once for a scalar and once for a vector.
4.2.4. Normalizing Vectors. Normalize each of the following vectors:
a). (1, -2, .5); b). (8, 6); c). (4, 3)

4.3. The Dot Product.

There are two other powerful tools that facilitate working with vectors: the dot (or inner) product, and the
cross product. The dot product produces a scalar; the cross product works only on three dimensional vectors
and produces another vector. In this section we review the basic properties of the dot product, principally to
develop the notion of perpendicularity. We then work with the dot product to solve a number of important
geometric problems in graphics. Then the cross product is introduced, and used to solve a number of 3D
geometric problems.
The dot product of two vectors is simple to define and compute. For two-dimensional vectors, (a1, a2) and
(b1, b2), it is simply the scalar whose value is a1b1 + a2b2. Thus to calculate it, multiply corresponding
components of the two vectors, and add the results. For example, the dot product of (3, 4) and (1, 6) is 27,
and that of (2, 3) and (9, -6) is 0.
The definition of the dot product generalizes easily to n dimensions:
Definition: The Dot Product
The dot product d of two n-dimensional vectors v = (v1, v2, . . . , vn) and w = (w1, w2, . . . , wn) is denoted as
v w and has the value
n

d = v w = vi wi

(4.11)

i =1

Example 4.3.1:
The dot product of (2, 3, 1) and (0, 4, -1) is 11.
(2, 2, 2, 2) (4, 1, 2, 1.1) = 16.2.
(1, 0, 1, 0, 1) (0, 1, 0, 1, 0) = 0.
(169, 0, 43) (0, 375.3, 0) = 0.

Hill - Chapter 4

09/23/99

page 8

4.3.1. Properties of the Dot Product

The dot product exhibits four major properties that we frequently exploit and that follow easily (see the
exercises) from its basic definition:
1
2
3
4

Symmetry:
Linearity:
Homogeneity:
|b|2 = b b

ab=ba
(a + c) b = a b + c b
(sa) b = s (a b)

The first states that the order in which the two vectors are combined does not matter: the dot product is
commutative. The next two proclaim that the dot product is linear; that is, the dot product of a sum of
vectors can be expressed as the sum of the individual dot products, and scaling a vector scales the value of
the dot product. The last property is also useful, as it asserts that taking the dot product of a vector with itself
yields the square of the length of the vector. It appears frequently in the form |b| = bb .
The following manipulations show how these properties can be used to simplify an expression involving dot
products. The result itself will be used in the next section.
Example 4.3.2: Simplification of |a - b|2.
Simplify the expression for the length (squared) of the difference of two vectors, a and b, to obtain the
following relation:
| a - b |2 = | a |2 - 2 a b + | b |2

(4.12)

The derivation proceeds as follows: Give the name C to the expression | a - b |2. By the fourth property, C is
the dot product:
C = | a - b |2 = (a - b) (a - b).
Using linearity: C = a (a - b) - b (a - b).
Using symmetry and linearity to simplify this further: C = a a - 2a b + b b.
Using the fourth property above to obtain C = | a |2 - 2 a b + | b |2 gives the desired result.
By replacing the minus with a plus in this relation, the following similar and useful relation emerges:
| a + b |2 = | a |2 + 2a b + | b |

(4.13)

4.3.2. The Angle Between Two Vectors.

The most important application of the dot product is in finding the angle between two vectors, or between
two intersecting lines. Figure 4.9 shows the 2D case, where vectors b and c lie at angles b, and c,
relative to the x-axis. Now from elementary trigonometry:
y
c

b
c

Figure 4.9. Finding the angle between two vectors.

Hill - Chapter 4

09/23/99

page 9

b = ( |b| cos b , |b| sin b)

c = ( |c| cos c , |c| sin c).
Thus their dot product is

b c = | b| | c|cos c cos b + | b| | c|sin b sin c

= b c cos( c b )
so we have, for any two vectors b and c:
b c = |b| |c| cos()

(4.14)

where is the angle from b to c. Thus b c varies as the cosine of the angle from b to c. The same result
holds for vectors of three, four, or any number of dimensions.
To obtain a slightly more compact form, divide through both sides by |b| |c| and use the unit vector notation

b = b / b to obtain
cos( ) = b c

(4.15)

This is the desired result: The cosine of the angle between two vectors b and c is the dot product of their
normalized versions.
Example 4.3.3. Find the angle between b = (3, 4) and c = (5, 2).

= (3 / 5, 4 / 5) and c = (.9285, .3714). The dot product

Solution: Form |b| = 5 and |c| = 5.385 so that b
b c = .85422 = cos(), so that = 31.326o. This can be checked by plotting the two vectors on graph paper
and measuring the angle between them.
4.3.3. The Sign of bc, and Perpendicularity.
Recall that cos() is positive if || is less than 90o, zero if || equals 90o, and negative if || exceeds 90o.
Because the dot product of two vectors is proportional to the cosine of the angle between them, we can
therefore observe immediately that two vectors (of any nonzero length) are

less than
exactly
more than

90o apart
90o apart
90o apart

if b c > 0;
if b c = 0;
if b c < 0;

(4.16)

This is indicated by Figure 4.10. The sign of the dot product is used in many algorithmic tests.
b

b
c
c

b c> 0

b c= 0

c
b c< 0

Figure 4.10. The sign of the dot product.

The case in which the vectors are 90o apart, or perpendicular, is of special importance.
Definition:
Vectors b and c are perpendicular if b c = 0.
(4.17)

Hill - Chapter 4

09/23/99

page 10

Other names for perpendicular are orthogonal and normal, and we shall use all three interchangeably.
The most familiar examples of orthogonal vectors are those aimed along the axes of 2D and 3D coordinate
systems, as shown in Figure 4.11. In part a) the 2D vectors (1, 0) and (0, 1) are mutually perpendicular unit
vectors. The 3D versions are so commonly used they are called the standard unit vectors and are given
names i, j, and k.
y
a).

b).

y
j

(0,1)

(1,0)

x
x

Figure 4.11. The standard unit vectors.

Definition:
The standard unit vectors in 3D have components:
and k = (0, 0, 1).
i = (1, 0, 0),
j = (0, 1, 0),

(4.18)

Part b) of the figure shows them for a right-handed system, and part c) shows them for a left-handed
system. Note that k always points in the positive z direction.
Using these definitions any 3D vector such as (a, b, c) can be written in the alternative form:
(a, b, c) = a i + b j + c k

(4.19)

Example 4.3.4. Notice that v = (2, 5, -1) is clearly the same as 2 (1, 0, 0) + 5 (0, 1, 0) -1 (0, 0, 1), which is
recognized as 2 i + 5 j - k.
This form presents a vector as a sum of separate elementary component vectors, so it simplifies various
pencil-and-paper calculations. It is particularly convenient when dealing with the cross product, discussed in
Section 4.4.
Practice Exercises.
4.3.1. Alternate proof of b c = |b| |c| cos . Note that b and c form two sides of a triangle, and the third
side is b - c. Use the law of cosines to obtain the square of the length of b - c in terms of the lengths of b and
c and the cosine of . Compare this with Equation 4.13 to obtain the desired result.
4.3.2. Find the Angle. Calculate the angle between the vectors (2, 3) and (-3, 1), and check the result
visually using graph paper. Then compute the angle between the 3D vectors (1, 3, -2) and (3, 3, 1).
4.3.3. Testing for Perpendicularity. Which pairs of the following vectors are perpendicular to one another:
(3, 4, 1), (2, 1, 1), (-3, -4, 1), (0, 0, 0), (1, -2, 0), (4, 4, 4), (0, -1, 4), and (2, 2, 1)?
4.3.4. Pythagorean Theorem. Refer to Equations 4.12 and 4.13. For the case in which a and b are
perpendicular, these expressions have the same value, which seems to make no sense geometrically. Show
that it works all right, and relate the result to the Pythagorean theorem.

4.3.4. The 2D Perp Vector.

Hill - Chapter 4

09/23/99

page 11

Suppose the 2D vector a has components (ax, ay). What vectors are perpendicular to it? One way to obtain
such a vector is to interchange the x- and y- components and negate one of them.3 Let b = (-ay, ax). Then
the dot product a b equals 0 so a and b are indeed perpendicular. For instance, if a = (4,7) then b = (-7, 4)
is a vector normal to a. There are infinitely many vectors normal to any a, since any scalar multiple of b,
such as (-21, 12) and (7, -4) is also normal to a. (Sketch several of them for a given a.)
It is convenient to have a symbol for one particular vector that is normal to a given 2D vector a. We use
the symbol (pronounced perp) for this.
Definition: Given a = (ax, ay), a = (-ay, ax) is the counterclockwise
(4.20)
perpendicular to a.
Note that a and a have the same length: |a| = |a| . Figure 4.12a shows an arbitrary vector a and the
resulting a . Note that moving from the a direction to direction a requires a left turn. (Making a right
turn is equivalent to turning in the direction -a .)
a).

b).
a

a
-a

Figure 4.12. The vector a perpendicular to a.

We show in the next section how this notation can be put to good use. Figure 4.12b shows that in three
dimensions no single vector lies in the direction perpendicular to a given 3D vector a, since any of the
vectors lying in the plane perpendicular to a will do. However, the cross product developed later will
provide a simple tool for dealing with such vectors.
Practice Exercises.

4.3.5. Some Pleasant Properties of a. It is useful in some discussions to view the perp symbol as an
operator that performs a rotate 90 left operation on its argument, so that a is the vector produced by
applying the to vector a, much as x is the value produced by applying the square root operator to x.
Viewing in this way, show that it enjoys the following properties:
a). Linearity: (a + b) = a + b and (Aa) = Aa for any scalar A;
b). a = (a) = -a
(two perps make a reversal)

4.3.6. The perp dot product. Interesting things happen when we dot the perp of a vector with another
vector, as in a. b. We call this the perp dot product [hill95]. Use the basic definition of a above to
show:

(value of the perp dot product)

a b = ax by aybx

a.a = 0,
|a|2 = |a|2.
a. b = - b . a,

(a is perpendicular to a)
(a and a have the same length)

(4.21)

(antisymmetric)

3 This is equivalent to the familiar fact that perpendicular lines have slopes that are negative reciprocals of one another.
In Chapter 5 we see the interchange and negate operation arise naturally in connection with a rotation of 90 degrees.

Hill - Chapter 4

09/23/99

page 12

The fourth fact shows that the perp dot product is antisymmetric: moving the from one vector to the
other reverses the sign of the dot product. Other useful properties of the perp dot product will be discussed
as they are needed.
4.3.7. Calculate one. Compute a b and a. b for a = (3,4) and b = (2,1).
4.3.8. Its a determinant. Show that a. b can be written as the determinant (for definitions of matrices
and determinants see Appendix 2):

a b =

ax
bx

ay
by

4.3.9. Other goodies.

a). Show that (a.b)2 + (a.b)2 = |a|2|b|2.
b). Show that if a + b + c = 0 then a b = b c = c a.

4.3.5. Orthogonal Projections, and the Distance from a Point to a Line.

Three geometric problems arise frequently in graphics applications: projecting a vector onto a given
vector, resolving a vector into its components in one direction and another, and finding the distance
between a point and a line. All three problems are simplified if we use the perp vector and the perp dot
product.
Figure 4.13a shows the basic ingredients. We are given two points A and C, and a vector v. These questions arise:
b).

a).

c).

C
v

c
A

c
F

Figure 4.13. Resolving a vector into two orthogonal vectors.

a). How far is the point C from the line L that passes through A in the direction v?
b). If we drop a perpendicular from C onto L, where does it hit L?
c). How do we decompose the vector c = C - A into a part along the line L and a part perpendicular to L?
Figure 4.13.b defines some additional quantities: v is the vector v rotated 90 degrees CCW. Dropping a
perpendicular from C onto line L we say that the vector c is resolved into the portion Kv along v and the
portion M v perpendicular to v, where K and M are some constants to be determined. Then we have
c = Kv + M v

(4.22)

Given c and v we want to solve for K and M. Once found, we say that the orthogonal projection of c onto
v is Kv, and that the distance from C to the line is |Mv|.
Figure 4.13c shows a situation where these questions might arise. We wish to analyze how the gravitational
force vector G acts on the block to pull it down the incline. To do this we must resolve G into the force F
acting along the incline and the force B acting perpendicular to the incline. That is, find F and B such that
G = F + B.
Equation 4.22 is really two equations: the left and right hand sides must agree for the x-components and
they also must agree for the y-components. There are two unknowns K and M. So we have two equations
in two unknowns, and Cramers rule can be applied. But who remembers Cramers rule? We use a trick

Hill - Chapter 4

09/23/99

page 13

here that is easy to remember and immediately reveals the solution. It is equivalent to Cramers rule, but
simpler to apply.
The trick in solving two equations in two unknowns is to eliminate one of the variables. We do this by
forming the dot product of both sides with the vector v:
cv = Kvv + M vv

(4.23)

Happily, the term vv vanishes, (why?), yielding K immediately:

c v
.
vv

Similarly dot both sides of Equation 4.3.12 with v to obtain M:

c v
v v

where we have used the third property in Equation 4.21. Putting these together we have

v c
v c
c = 2 v+
v
| v |
| v |2

(resolving c into v and v)

(4.24)

This equality holds for any vectors c and v. The part along v is known as the orthogonal projection of c
onto the vector v. The second term gives the difference term explicitly and compactly. Its size is the
distance from C to the line:

v c
v c
v =
,
distance =
| v |2
v
(Check that the second form really equals the first). Referring to Figure 4.13b we can say: the distance
from a point C to the line through A in the direction v is:

distance =

v (C A)
v

(4.25)

Example 4.3.5. Find the orthogonal projection of the vector c = (6, 4) onto a = (1, 2). (Sketch the relevant
vectors.) Solution: Evaluate the first term in Equation 4.24, obtaining the vector (14, 28) / 5.
Example 4.3.6: How far is the point C = (6,4) from the line that passes through (1,1) and (4,9)? Solution:
Set A = (1, 1), use v = (4, 9) - (1, 1) = (3, 8), and evaluate distance in Equation 4.25. The result is:
d = 31 / 73 .
Practice Exercises.
4.3.10. Resolve it. Express vector g = (4, 7) as a linear combination of b = (3, 5) and b. How far is (4, 2)
+ g from the line through (4, 2) that moves in the direction b?
4.3.11. A Block pulled down an incline. A block rests on an incline tilted 30 from the horizontal. Gravity
exerts a force of one newton on the block. What is the force that is trying to move the block along the
incline?
4.3.12. How far is it? How far from the line through (2, 5) and ( 4, -1) does the point (6, 11) lie? Check your
result on graph paper.

Hill - Chapter 4

09/23/99

page 14

4.3.6. Applications of Projection: Reflections.

To display the reflection of light from a mirror, or the behavior of billiard balls bouncing off one another,
we need to find the direction that an object takes upon being reflected at a given surface. In a case study at
the end of this chapter we describe an application to trace a ray of light as it bounces around inside a
reflective chamber, or a billiard ball as it bounces around a pool table. At each bounce a reflection is made
to a new direction, as derived in this section.
When light reflects from a mirror we know that the angle of reflection must equal the angle of incidence.
We next show how to use vectors and projections to compute this new direction. We can think in terms of
two-dimensional vectors for simplicity, but because the derivation does not explicitly state the dimension of
the vectors involved, the same result applies in three dimensions for reflections from a surface.
Figure 4.14a shows a ray having direction a, hitting line L, and reflecting in (as yet unknown) direction r.
The vector n is perpendicular to the line. Angle 1 in the figure must equal angle 2. How is r related to a
and n? Figure 4.14b shows a resolved into the portion m along n and the portion e orthogonal to n.
Because of symmetry, r has the same component e orthogonal to n, but the opposite component along n,
and so r = e - m. Because e = a - m, this gives r = a - 2m. Now m is the orthogonal projection of a onto n,
so by Equation 4.24 m is

a).

b).

1 2
L

-m

4.14. Reflection of a ray from a surface.

an

2 n = (a n )n
|n|

(4.26)

is the unit length version of n) and so we obtain the result

(recall n
r = a 2(a n )n

(direction of the reflected ray)

(4.27)

In three dimensions physics demands that the reflected direction r must lie in the plane defined by n and a.
The expression for r above indeed supports this, as we show in Chapter five.
Example 4.3.7. Let a = (4, -2) and n = (0, 3). Then Equation 4.27 yields r = (4, 2), as expected. Both the
angle of incidence and reflection are equal to tan-1(2).
Practice Exercises.
4.3.13. Find the Reflected Direction. For a = (2, 3) and n = (-2, 1), find the direction of the reflection.
4.3.14. Lengths of the Incident and Reflected Vectors. Using Equation 4.27 and properties of the dot
product, show that |r| = |a|.

4.4. The Cross Product of Two Vectors.

The cross product (also called the vector product) of two vectors is another vector. It has many useful
properties, but the one we use most often is that it is perpendicular to both of the given vectors. The cross
product is defined only for three-dimensional vectors.

Hill - Chapter 4

09/23/99

page 15

Given the 3D vectors a = (ax, ay, az) and b = (bx, by, bz), their cross product is denoted as a b . It is
defined in terms of the standard unit vectors i, j, and k (see Equation 4.18) by
Definition of a b :

a b = (ay bz az by )i + (azbx ax bz ) j + (a xby ay bx )k

(4.28)

(It can actually be derived from more fundamental principles: See the exercises.) As this form is rather
difficult to remember, it is often written as an easily remembered determinant (see Appendix 2 for a review
of determinants).

a b = ax
bx

ay
by

az
bz

(4.29)

Remembering how to form the cross product thus requires only remembering how to form a determinant.
Example 4.4.1. For a = (3, 0, 2) and b = (4, 1, 8), direct calculation shows that a b = -2i -16j + 3k. What
is b a?
From this definition one can easily show the following algebraic properties of the cross product:

i j= k
jk =i
ki = j

2. a b = b a
3. a (b + c ) = a b + a c
4. (sa) b = s(a b)

(antisymmetry)
(linearity)
(homogeneity)

(4.30)

These equations are true in both left-handed and right-handed coordinate systems. Note the logical
(alphabetical) ordering of ingredients in the equation i j = k , which also provides a handy mnemonic
device for remembering the direction of cross products.
Practice Exercises.
4.4.1. Demonstrate the Four Properties. Prove each of the preceding four properties given for the cross
product.
4.4.2. Derivation of the Cross Product. The form in Equation 4.28, presented as a definition, can actually
be derived from more fundamental ideas. We need only assume that:
a. The cross product operation is linear.
b. The cross product of a vector with itself is 0.
c. i j = k , j k = i , and k i = j .
By writing a = ax i + ay j + az k and b = bx i + by j + bz k, apply these rules to derive the proper form for
a b.
4.4.3. Is a b perpendicular to a? Show that the cross product of vectors a and b is indeed
perpendicular to a.
4.4.4. Vector Products. Find the vector b = (bx, by, bz) that satisfies the cross product relation a b = c,
where a = (2, 1, 3) and c = (2, -4, 0). Is there only one such vector?

Hill - Chapter 4

09/23/99

page 16

4.4.5. Nonassociativity of the Cross Product. Show that the cross product is not associative. That is, that
a (b c) is not necessarily the same as (a b) c .
4.4.6. Another Useful Fact. Show by direct calculation on the components that the length of the cross
product has the form:

ab =

2
a b (a b)
2

4.4.1. Geometric Interpretation of the Cross Product.

By definition the cross product a b of two vectors is another vector, but how is it related geometrically to the
others, and why is it of interest? Figure 4.15 gives the answer. The cross product a b has the following useful
properties (whose proofs are requested in the exercises):
ab
Area = | a b |
a

Figure 4.15. Interpretation of the cross product.

1. a b is perpendicular (orthogonal) to both a and b.
2. The length of a b equals the area of the parallelogram determined by a and b. This area is equal to

|a b|=|a||b|sin( )

(4.31)

where is the angle between a and b, measured from a to b or b to a, whichever produces an angle less than 180
degrees. As a special case, a b = 0 if, and only if, a and b have the same or opposite directions or if either has
zero length. What is the magnitude of the cross product if a and b are perpendicular?
3. The sense of a b is given by the right-hand rule when working in a right-handed system. For example,
twist the fingers of your right hand from a to b, and then a b will point in the direction of your thumb. (When
working in a left-handed system, use your left hand instead.) Note that i j = k supports this.
Example 4.4.2. Let a = (1, 0, 1) and b = (1, 0, 0). These vectors are easy to visualize, as they both lie in the x, zplane. (Sketch them.) The area of the parallelogram defined by a and b is easily seen to be 1. Because a b is
orthogonal to both a and b, we expect it to be parallel to the y-axis and hence be proportional to j. In either a
right-handed or a left-handed system, sweeping the fingers of the appropriate hand from a to b reveals a thumb
pointed along the positive y-axis. Direct calculation based on Equation 4.28 confirms all of this: a b = j.
Practice Exercise 4.4.7. Proving the Properties. Prove the three properties given above for the cross product.

4.4.2. Finding the Normal to a Plane.

As we shall see in the next section, we sometimes must compute the components of the normal vector n to a
plane.
If the plane is known to pass through three specific points, the cross product provides the tool to accomplish this.
Any three points, P1, P2, P3, determine a unique plane, as long as the points dont lie in a straight line. Figure
4.16 shows this situation.

Hill - Chapter 4

09/23/99

page 17

P1
a
P2
x

z
a xb

Figure 4.16. Finding the plane through three given points.

To find the normal vector, build two vectors, a = P2 - P1 and b = P3 - P1. Their cross product, n = a b ,
must be normal to both a and b, so it is normal to every line in the plane (why?). It is therefore the desired
normal vector. (What happens if the three points do lie in a straight line?) Any scalar multiple of this cross
product is also a normal vector, including b a , which points in the opposite direction.
Example 4.4.3. Find the normal vector to the plane that passes through the points (1, 0, 2), (2, 3, 0), and (1, 2, 4).
Solution: By direct calculation, a = (2, 3, 0) - (1, 0, 2) = (1, 3, -2), and b = (1, 2, 4) - (1, 0, 2) = (0, 2, 2), and so
their cross product n = (10, -2, 2).
Note: Since a cross product involves the subtraction of various quantities (see Equation 4.28), this method for
finding n is vulnerable to numerical inaccuracies, especially when the angle between a and b is small. We
develop a more robust method later for finding normal vectors in practice.
Practice Exercises.
4.4.8. Does the choice of points matter? Is the same plane obtained as in Example 4.4.3 if we use the points in
a different order, say, a = (1, 0, 2) - (2, 3, 0) and b = (1, 2, 4) - (2, 3, 0)? Show that the same plane does result.
4.4.9. Finding Some Planes. For each of the following triplets of points, find the normal vector to the plane (if it
exists) that passes through the triplet.
a. P1 = (1, 1, 1), P2 = (1, 2, 1), P3 = (3, 0, 4)
b. P1 = (8, 9, 7), P2 = ( -8, -9, -7), P3 = (1, 2, 1)
c. P1 = (6, 3, -4), P2 = (0, 0, 0), P3 = (2, 1, -1)
d. P1 = (0, 0, 0), P2 = (1, 1, 1), P3 = (2, 2, 2).
4.4.10. Finding the normal vectors. Calculate the normal vectors to each of the faces of the two objects shown
in Figure 4.17. The cube has vertices (1,1,1) and the tetrahedron has vertices (0,0,0), (0,0,1), (1,0,0), and
(0,1,0).
a).

b).

Figure 4.17. Finding the normal vectors to faces.

4.5. Representations of Key Geometric Objects.

In the preceding sections we have discussed some basic ideas of vectors and their application to important
geometric problems that arise in graphics. Now we develop the fundamental ideas that facilitate working

Hill - Chapter 4

09/23/99

page 18

with lines and planes, which are central to graphics, and whose straightness and flatness makes them
easy to represent and manipulate.
What does it mean to represent a line or plane, and why is it important? The goal is to come up with a
formula or equation that distinguishes points that lie on the line from those that dont. This might be an
equation that is satisfied by all points on the line, and only those points. Or it might be a function that returns
different points in the line as some parameter is varied. The representation allows one to test such things as:
is point P on the line?, or where does the line intersect another line or some other object. Very importantly, a
line lying in a plane divides the plane into two parts, and we often need to ask whether point P lies on one
side or the other of the line.
In order to deal properly with lines and planes we must, somewhat unexpectedly, go back to basics and
review how points and vectors differ, and how each is represented. The need for this arises because, to
represent a line or plane we must add points together, and scale points, operations that for points are
nonsensical. To see what is really going on we introduce the notion of a coordinate frame, that makes clear
the significant difference between a point and a vector, and reveals in what sense it is legitimate to add
points. The use of coordinate frames leads ultimately to the notion of homogeneous coordinates, which
is a central tool in computer graphics, and greatly simplifies many algorithms. We will make explicit use of
coordinate frames in only a few places in the book, most notably when changing coordinate systems and
flying cameras around a scene (see Chapters 5, 6, and 7) 4. But even when not explicitly mentioned, an
underlying coordinate frame will be present in every situation.

4.5.1. Coordinate Systems and Coordinate Frames.

One doesnt discover new lands without consenting
to lose sight of the shore for a very long time.
Andre Gide
When discussing vectors in previous sections we say, for instance, that a vector v = (3, 2, 7), meaning it is a
certain 3-tuple. We say the same for a point, as in point P = (5, 3, 1). This makes it seem that points and
vectors are the same thing. But points and vectors are very different creatures: points have location but no
size or direction; vectors have size and direction but no location.
What we mean by v = (3, 2, 7), of course, is that the vector v has components (3, 2, 7) in the underlying
coordinate system. Similarly, P = (5, 3, 1) means point P has coordinates (5, 3, 1) in the underlying
coordinate system. Normally this confusion between the object and its representation presents no problem.
The problem arises when there is more than one coordinate system (a very common occurrence in graphics),
and when you transform points or vectors from one system into another.
We usually think of a coordinate system as three axes emanating from an origin, as in Figure 4.2b. But in
fact a coordinate system is located somewhere in the world, and its axes are best described by three
vectors that point in mutually perpendicular directions. In particular it is important to make explicit the
location of the coordinate system. So we extend the notion of a 3D coordinate system5 to that of a 3D
coordinate frame. A coordinate frame consists of a specific point, , called the origin, and three
mutually perpendicular unit vectors6, a, b, and c. Figure 4.18 shows a coordinate frame residing at some
point within the world, with its vectors a, b, and c drawn so they appear to emanate from like axes.

4 This is an area where graphics programmers can easily go astray: their programs produce pictures that look OK for simple
situations, and become mysteriously and glaringly wrong when things get more complex.
5 The ideas for a 2D system are essentially identical.
6 In more general contexts the vectors need not be mutually perpendicular, but rather only linearly independent (such that, roughly,
none of them is a linear combination of the other two). The coordinate frames we work with will always have perpendicular axis
vectors.

Hill - Chapter 4

09/23/99

page 19

Figure 4.18. A coordinate frame positioned in the world.

Now to represent a vector v we find three numbers, (v1, v2, v3) such that
v = v1a + v2 b + v3 c

(4.32)

and say that v has the representation (v1, v2, v3) in this system.
On the other hand, to represent a point, P, we view its location as an offset from the origin by a certain
amount: we represent the vector P - by finding three numbers (p1, p2, p3) such that:
P - = p1a + p2b + p3c
and then equivalently write P itself as:
P = + p1a + p2b + p3c

(4.33)

The representation of P is not just a 3-tuple, but a 3-tuple along with an origin. P is at a location that is
offset from the origin by p1a + p2b + p3c. The basic idea is to make the origin of the coordinate system
explicit. This becomes important only when there is more than one coordinate frame, and when transforming
one frame into another.
Note that when we earlier defined the standard unit vectors i, j, and k as (1, 0, 0), (0, 1, 0), and (0, 0, 1),
respectively, we were actually defining their representations in an underlying coordinate frame. Since by
Equation 4.32 i = 1a + 0b + 0c, vector i is actually just a itself! Its a matter of naming: whether you are
talking about the vector or about its representation in a coordinate frame. We usually dont bother to
distinguish them.
Note that you cant explicitly say where is, or cite the directions of a, b, and c: To do so requires having
some other coordinate frame in which to represent this one. In terms of its own coordinate frame, has the
representation (0, 0, 0), a has the representation (1, 0, 0), etc.
The homogeneous representation of a point and a vector.
It is useful to represent both points and vectors using the same set of basic underlying objects, (a, b, c, ).
From Equations 4.32 and 4.33 the vector v = v1a + v2 b + v3 c then needs the four coefficients (v1, v2, v3, 0)
whereas the point P = p1a + p2b + p3c + needs the four coefficients (p1, p2, p3, 1). The fourth component
designates whether the object does or does not include . We can formally write any v and P using a matrix
multiplication (multiplying a row vector by a column vector - see Appendix 2):

Hill - Chapter 4

09/23/99

page 20

v
v
v = (a, b, c, )
v
0
p
p
P = (a, b, c, )
p
1
1
2

(4.34)

1
2

(4.35)

Here the row matrix captures the nature of the coordinate frame, and the column vector captures the
representation of the specific object of interest. Thus vectors and points have different representations: there
is a fourth component of 0 for a vector and 1 for a point. This is often called the homogeneous
representation.7 The use of homogeneous coordinates is one of the hallmarks of computer graphics, as it
helps to keep straight the distinction between points and vectors, and provides a compact notation when
working with affine transformations. It pays off in a computer program to represent the points and vectors of
interest in homogeneous coordinates as 4-tuples, by appending a 1 or 08. This is particularly true when we
must convert between one coordinate frame and another in which points and vectors are represented.
It is simple to convert between the ordinary representation of a point or vector (a 3-tuple for 3D objects or
a 2-tuple for 2D objects) and the homogeneous form:
To go from ordinary to homogeneous coordinates:
if its a point append a 1;
if its a vector, append a 0;
To go from homogeneous coordinates to ordinary coordinates:
If its a vector its final coordinate is 0. Delete the 0.
If its a point its final coordinate is 1 Delete the 1.
OpenGL uses 4D homogeneous coordinates for all its vertices. If you send it a 3-tuple in the form (x, y, z), it
converts it immediately to (x, y, z, 1). If you send it a 2D point (x, y), it first appends a 0 for the z-component
and then a 1, to form (x, y, 0, 1). All computations are done within OpenGL in 4D homogeneous
coordinates.
Linear Combinations of Vectors .
Note how nicely some things work out in homogeneous coordinates when we combine vectors coordinatewise: all the definitions and manipulations are consistent:
The difference of two points (x, y, z, 1) and ( u, v, w, 1) is ( x - u, y - v, z - w, 0), which is, as expected, a
vector.
The sum of a point (x, y, z, 1) and a vector (d, e, f, 0) is (x + d, y + e, z + f, 1), another point;
Two vectors can be added: (d, e, f, 0) + (m, n, r, 0) = (d + m, e + n, f + r, 0) which produces another vector;
It is meaningful to scale a vector: 3(d, e, f, 0) = (3d, 3e, 3f, 0);

7 Actually we are only going part of the way in this discussion. As we see in Chapter 7 when studying projections, homogeneous
coordinates in that context permit an additional operation, which makes them truly homogeneous. Until we examine projections this
operation need not be introduced.
8 In the 2D case, points are 3-tuples (p , p , 1) and vectors are 3-tuples (v , v , 0).
1
2
1
2

Hill - Chapter 4

09/23/99

page 21

It is meaningful to form any linear combination of vectors. Let the vectors be v = (v1 , v2, v3, 0) and w =
(w1, w2, w3, 0). Then using arbitrary scalars a and b, we form av + bw = (av1 + bw1, av2 + bw2, av3 + bw3, 0),
which is a legitimate vector.
Forming a linear combination of vectors is well defined, but does it make sense for points? The answer is no,
except in one special case, as we explore next.

4.5.2. Affine Combinations of Points.

Consider forming a linear combination of two points, P = (P1, P2, P3,1) and R = (R1, R2, R3, 1) , using the
scalars f and g:
fP + gR = (fP1 + gR1, fP2 + gR2, fP3 + gR3, f + g)
We know this is a legitimate vector if f + g = 0 (why?). But we shall see that it is not a legitimate point
unless f + g = 1! Recall from Equation 4.2 that when the coefficients of a linear combination sum to 1 it is
called an affine combination. So we see that the only linear combination of points that is legitimate is an
affine combination. Thus, for example, the object 0.3P + 0.7R is a legitimate point, as are 2.7P - 1.7R and
the midpoint 0.5P + 0.5R, but P + R is not a point. For three points, P, R, and Q we can form the legal point
0.3P + 0.9R - 0.2Q, but not P + Q - 0.9R..
Fact: any affine combination of points is a legitimate point.
But whats wrong geometrically with forming any linear combination of two points, say
E = fP + gR

(4.36)

when f + g is different from 1? The problem arises if we shift the origin of the coordinate system
[Goldman85]. Suppose the origin is shifted by vector u, so that P is altered to P + u and R is shifted to R +
u. If E is a legitimate point, it too must be shifted to the new point E = E + u. But instead we have
E = fP + gR + (f + g)u
which is not E + u unless f + g = 1.
The failure of a simple sum P1 + P2 of two points to be a true point is shown in Figure 4.19. Points P1 and P2
are shown represented in two coordinate systems, one offset from the other. Viewing each point as the head
of a vector bound to its origin, we see that the sum P1 + P2 yields two different points in the two systems.
Therefore P1 + P2 depends on the choice of coordinate system. Note, by way of contrast, that the affine
combination 0.5(P1 + P2) does not depend on this choice.
P1
depends on
system

System 2

(P1+ P 2)

P1
System 1
(P1+ P 2)/2
P1

Figure 4.19. Adding points is not legal.

A Point plus a Vector is an Affine Combination of Points.

Hill - Chapter 4

09/23/99

page 22

There is another way of examining affine sums of points that is interesting on its own, and also leads to a
useful tool in graphics. It doesnt require the use of homogeneous coordinates.
Consider forming a point as a point A offset by a vector v that has been scaled by scalar t: A + tv. This is the
sum of a point and a vector so it is a legitimate point. If we take as vector v the difference between some
other point B and A: v = B - A then we have the point P:
P = A + t(B - A)

(4.37)

which is also a legitimate point. But now rewrite it algebraically as:

P = tB + (1 - t)A

(4.38)

and it is seen to be an affine combination of points (why?). This further legitimizes writing affine sums of
points. In fact, any affine sum of points can be written as a point plus a vector (see the exercises). If you are
ever uncomfortable writing an affine sum of points as in Equation 4.38 (a form we will use often), simply
understand that it means the point given by Equation 4.37.
Example 4.5.1: The centroid of a triangle. Consider the triangle T with vertices A, B, and C shown in
Figure 4.20. We use the ideas above to show that the three medians of T meet at a point that lies 2/3 of the
way along each median. This is the centroid (center of gravity9) of T.
E
G
F
centroid C
D

Figure 4.20. The centroid of a triangle as an affine combination.

By definition the median from D is the line from D to the midpoint of the opposite side. Thus G = (E + F)/2.
We first ask where the point that is 2/3 of the way from D to G lies? Using the parametric form the desired
point must be D + (G - D)t with t = 2/3, which yields the affine combination C given by

D+E+ F
3

(Try it!) Heres the cute part [pedoe70]. Since this result is symmetrical in D, E, and F, it must also be 2/3 of
the way along the median from E, and 2/3 of the way along the median from F. Hence the 3 medians meet
there, and C is the centroid.
This result generalizes nicely for a regular polygon of N sides: the centroid is simply the average of the N
vertex locations, another affine combination. For an arbitrary polygon the formula is more complex
Practice Exercises.
4.5.1. Any affine combination of points is legitimate. Consider three scalars a, b, and c that sum to one,
and three points A, B, and C. The affine combination a A + b B + c C is a legal point because using c = 1 - a
- b it is seen to be the same as a A + b B + (1 - a - b) C = C + a (A - C) + b (B - C), the sum of a point and
two vectors (check this out!). To generalize: Given the affine combination of points w1A1 + w2A2 + ... +
wnAn, where w1 + w2 + ... + wn = 1, show that it can be written as a point plus a vector, and is therefore a
legitimate point.
4.5.2. Shifting the coordinate system [Goldman85]. Consider the general situation of forming a linear
combination of m points:
9The reference to gravity arises because if a thin plate is cut in the shape of T, the plate hangs level if suspended by a
thread attached at the centroid. Gravity pulls equally on all sides of the centroid, so the plate is balanced.

Hill - Chapter 4

09/23/99

page 23

E = ai Pi
i =1

We ask whether E is a point, a vector, or nothing at all? By considering the effect of a shift in each Pi by u
show that E is shifted to E = E + S u, where S is the sum of the coefficients:
m

S = ai
i =1

Show that:
i). E is a point if S = 1.
ii). E is a vector if S = 0.
iii). E is meaningless for other values of S.

4.5.3. Linear Interpolation of two points.

The affine combination of points expressed in Equation 4.33:
P = A(1 - t) + Bt
performs linear interpolation between the points A and B. That is, the x-component Px(t)
provides a value that is fraction t of the way between the value Ax and Bx, and similarly for the
y-component (and in 3D the z-component). This is a sufficiently important operation to warrant
a name, and lerp() (for linear interpolation) has become popular. In one dimension, lerp(a, b,
t) provides a number that is the fraction t of the way from a to b. Figure 4.21 provides a simple
implementation of lerp().
float lerp(float a, float b, float t)
{
return a + (b - a) * t; // return a float
}
Figure 4.21. Linear interpolation effected by lerp().
Similarly, one often wants to compute the point P(t) that is fraction t of the way along the
straight line from point A to point B. This point is often called the tween (for in-between)
at t of points A and B. Each component of the resulting point is formed as the lerp() of the
corresponding components of A and B. A procedure
Point2 canvas:: Tween(Point2 A, Point2 B, float t) // tween A and B
is easily written (how?) to implement tweening. A 3D version is almost the same.
Example 4.5.2. Let A = (4, 9) and B = (3, 7). Then Tween(A, B, t) returns the point (4 - t, 9 - 2t), so
that Tween(A, B, 0.4) returns (3.6, 8.1). (Check this on graph paper.)

4.5.3. Tweening for Art and Animation.

Interesting animations can be created that show one figure being tweened into another. Its simplest if the
two figures are polylines (or families of polylines) based on the same number of points. Suppose the first
figure, A, is based on the polyline with points Ai , and the second polyline, B, is based on points Bi, for i =
0, . . . , n-1. We can form the polyline P(t), called the tween at t, by forming the points:
Pi(t) = (1 - t) Ai+t Bi
If we look at a succession of values for t between 0 and 1, say, t = 0, 0.1, 0.2, ..., 0.9, 1.0, we see that this
polyline begins with the shape of A and ends with the shape of B, but in between it is a blend of the two

Hill - Chapter 4

09/23/99

page 24

shapes. For small values of t it looks like A, but as t increases it warps (smoothly) towards a shape close to
B. For t = 0.25, for instance, point Pi(.25) of the tween is 25% of the way from A to B.
Figure 4.22 shows a simple example, in which polyline A has the shape of a house, and polyline B has the
shape of the letter T. The point R on the house corresponds to point S on the T. The various tweens of
point R on the house and point S on the T lie on the line between R and S. The tween for t = 1/2 lies at the
midpoint of RS. The in between polylines show the shapes of the tweens for t = 0, 0.25, 0.5, 0.75, and 1.0.
S
R

Figure 4.22. Tweening a "T" into a house.

Figure 4.23 shows drawTween(), that draws a tween of two polylines A and B, each having n vertices, at
the specified value of t.
void canvas:: drawTween(Point2 A[], Point2 B[], int n, float
t)
{
// draw the tween at time t between polylines A and B
for(int i = 0; i < n; i++)
{
Point2 P;
P = Tween(A[i], B[i],t);
if(i == 0)
moveTo(P.x, P.y);
else
lineTo(P.x, P.y);
}
}
Figure 4.23. Tweening two Polylines.
drawTween() could be used in an animation loop that tweens A and B back and forth, first as t increases
from 0 to 1, then as t decreases back to 0, etc. Double buffering, as discussed in Chapter 3, is used to make
the transition from one displayed tween to the next instantaneous.
for(t = 0.0, delT = 0.1; ; t += delT) // tween back and forth forever
{
<clear the buffer>
drawTween(A, B, n, t);
glutSwapBuffers();
if( t >= 1.0 || t <= 0.0) delT = - delT; // reverse the flow of t
}
Figure 4.24 shows an artistic use of this technique based on two sets of polylines. Three tweens are shown
(what values of t are used?). Because the two sets of polylines are drawn sufficiently far apart, there is
room to draw the tweens between them with no overlap, so that all five pictures fit nicely on one frame.
see Figure 7.11 from first edition
Figure 4.24. From man to woman. (Courtesy of Marc Infield.)
Susan E. Brennan of Hewlett Packard in Palo Alto, California, has produced caricatures of famous figures
using this method (see [dewdney88]). Figure 4.25 shows an example. The second and fourth faces are
based on digitized points for Elizabeth Taylor and John F. Kennedy. The third face is a tween, and the other
three are based on extrapolation. That is, values of t larger than 1 are used, so that the term (1 - t) is
negative. Extrapolation can produce caricature-like distortions, in some sense going to the other side of
polyline B from polyline A. Values of t less than 0 may also be used, with a similar effect.
see Figure 7.12 from 1st edition: Elizabeth Taylor to J.F. Kennedy

Hill - Chapter 4

09/23/99

page 25

Figure 4.25. Face Caricature: Tweening and extrapolation. (Courtesy of Susan Brennan.)
Tweening is used in the film industry to reduce the cost of producing animations such as cartoons. In earlier
days an artist had to draw 24 pictures for each second of film, because movies display 24 frames per
second. With the assistance of a computer, however, an artist need draw only the first and final pictures,
called key-frames, in certain sequences and let the others be generated automatically. For instance, if the
characters are not moving too rapidly in a certain one-half-second portion of a cartoon, the artist can draw
and digitize the first and final frames of this portion, and the computer can create 10 tweens using linear
interpolation, thereby saving a great deal of the artist's time. See the case study at the end of this chapter for
a programming project that produces these effects.
Practice Exercises.
4.5.3. A Limiting Case of Tweening. What is the effect of tweening when all of the points Ai in polyline A
are the same? How is polyline B distorted in its appearance in each tween?
4.5.4. An Extrapolation. Polyline A is a square with vertices (1, 1), (-1, 1), (-1, -1), (1, -1), and polyline B
is a wedge with vertices (4, 3), (5, -2), (4, 0), (3, -2). Sketch (by hand) the shape P(t) for t = -1, -0.5, 0.5,
and 1.5.
4.5.5. Extrapolation Versus Tweening. Suppose that five polyline pictures are displayed side by side.
From careful measurement you determine that the middle three are in-betweens of the first and the last, and
you calculate the values of t used. But someone claims that the last is actually an extrapolation of the first
and the fourth. Is there any way to tell whether this is true? If it is an extrapolation, can the value of t used
be determined? If so, what is it?

4.5.4. Preview: Quadratic and cubic tweening, and Bezier Curves.

In Chapter 8 we address the problem of designing complex shapes called Bezier curves. It is interesting to
note here that the underlying idea is simply tweening between a collection of points. With linear
interpolation above we partition unity into the pieces (1 - t) and t, and use these pieces to weight the
points A and B. We can extend this to quadratic interpolation by partitioning unity into three pieces. Just
rewrite 1 as
1 = ((1-t) + t)2
and expand it to produce the three pieces (1 - t)2, 2(1 - t) t, and t2. They obviously sum to one, so they can
be used to form the affine combination of points A, B, and C:
P(t) = (1 - t)2 A + 2(1 - t) B + t C
This is the Bezier curve for the points A, B, and C. Figure 4.26a shows the shape of P(t) as t varies from 0
to 1. It flows smoothly from A to C. (Notice that the curve misses the middle point.) Going further, one can
expand ((1 - t) + t )3 into four pieces (which ones?) which can be used to do cubic interpolation between
four points A, B, C and D, as shown in Figure 4.26b.
b).

a).
B

P(0)

P(t)

P(0)
P(1)

A
P(1)

Figure 4.26. Bezier curves as Tweening.

Hill - Chapter 4

09/23/99

page 26

Practice Exercise 4.5.6. Try it out. Draw three points A, B, and C on a piece of graph paper. For each of
the values t = 0, .1, .2, ..., .9, 1 compute the position of P(t) in Equation 4.38, and draw the polyline that
passes through these points. Is it always a parabola?

4.5.5. Representing Lines and Planes.

We now turn to developing the principal forms in which lines and planes are represented mathematically. It
is quite common to find data structures within a graphics program that capture a line or plane using one of
these forms.

Lines in 2D and 3D space.

A line is defined by two points, say C and B (see Figure 4.27a). It is infinite in length, passing through the
points and extending forever in both directions. A line segment (segment for short) is also defined by two
points, its endpoints, but extends only from one endpoint to the other (Figure 4.27b). Its parent line is the
infinite line that passes through its endpoints. A ray is semi-infinite. It is specified by a point and a
direction. It starts at a point and extends infinitely far in a given direction (Figure 4.27c).
c). r a y
b ). li n e s e g m en t
a ). lin e
B
B
B
C

C
s ta rti n g
p o in t

Figure 4.27. Lines, segments, and rays.

These objects are very familiar, yet it is useful to collect their important representations and properties in
one spot. We also describe the most important representation of all for a line in computer graphics, the
parametric representation.
The parametric representation of a line.
The construction in Equations 4.32 and 4.33 is very useful, because as t varies the point P traces out all of
the points on the straight line defined by C and B. The construction therefore gives us a way to name and
compute any point along this line.
This is done using a parameter t that distinguishes one point on the line from another. Call the line L, and
give the name L(t) to the position associated with t. Using b = B - C we have:
L(t) = C + b t

(4.39)

As t varies so does the position of L(t) along the line. (One often thinks of t as time, and uses language
such as: at time 0 ..., as time goes on.., or later to describe different parts of the line.) Figure 4.28
shows vector b and the line L passing through C and B. (A 2D version is shown but the 3D version uses the
same ideas.) Note where L(t) is located for various values of t. If t = 0, L(0) evaluates to C so at t = 0 we are
at point C. At t = l then L(1) = C + (B - C) = B. As t varies we add a longer or shorter version of b to the
point C, resulting in a new point along the line. If t is larger than 1 this point lies somewhere on the opposite
side of C from B, and when t is less than 0 it lies on the side of C opposite from B.

Hill - Chapter 4

09/23/99

page 27

@t > 1
B

y
v

@t = 1

L
@t = 0
@t < 0

x
Figure 4.28. Parametric representation L(t) of a line.
For a fixed value of t, say t = 0.6, Equation 4.39 gives a formula for exactly one point along the line through
C and B: the particular point L(0.6). Thus it is a description of a point. But since one can view it as a
function of t that generates the coordinates of every point on L as t varies, it is called the parametric
representation of line L.
The line, ray, and segment of Figure 4.26 are all represented by the same L(t) of Equation 4.39. They differ
parametrically only in the values of t that are relevant:
segment : 0 t 1
ray: 0 t <
line : - < t <

(4.40)

The ray starts at C when t = 0 and passes through B at t = 1, then continues forever as t increases. C is
often called the starting point of the ray.
A very useful fact is that L(t) lies fraction t of the way between C and B when t lies between 0 and 1. For
instance, when t = 1/2 the point L(0.5) is the midpoint between C and B, and when t = 0.3 the point L(0.3) is
30% of the way from C to B. This is clear from Equation 4.39 since |L(t) - C = |b| |t| and |B - C| = |b|, so the
value of |t| is the ratio of the distances |L(t) - C| to |B - C|, as claimed.
One can also speak of the speed with which the point L(t) moves along line L. Since it covers distance
|b| t in time t it is moving at constant speed |b|.
Example 4.5.2. A line in 2D. Find a parametric form for the line that passes through C= (3, 5) and B = (2,
7). Solution: Build vector b = B - C = (-1, 2) to obtain the parametric form L(t) = (3 - t, 2 + 2 t).
Example 4.5.3. A line in 3D. Find a parametric form for the line that passes through C= (3, 5,6) and B = (2,
7,3). Solution: Build vector b = B - C = (-1, 2, -3) to obtain the parametric form L(t) = (3 - t, 2 + 2 t, 6 - 3t).
Other parametrizations for a straight line are possible, although they are rarely used. For instance, the point
W(t) given by
W(t) = C + bt3
also sweeps over every point on L. It lies at C when t = 0 , and reaches B when t = 1. Unlike L(t), however,
W(t) accelerates along its path from C to B.
Point normal form for a line (the implicit form).
This is the same as the equation for a line, but we rewrite it in a way that better reveals the underlying
geometry. The familiar equation of a line in 2D has the form
fx+gy=1

Hill - Chapter 4

(4.41)

09/23/99

page 28

where f and g are some constants. The notion is that every point (x, y) that satisfies this equation lies on the
line, so it provides a condition for a point to be on the line. Note: This is true only for a line in 2D; a line in
3D requires two equations. So, unlike the parametric form that works perfectly well in 2D and 3D, the point
normal form only applies to lines in 2D.
This equation can be written using a dot product: (f, g) (x, y) = 1, so for every point on a line a certain dot
product must have the same value. We examine the geometric interpretation of the vector (f, g), and in so
doing develop the point normal form of a line. It is very useful in such tasks as clipping, hidden line
elimination, and ray tracing. Formally the point normal form makes no mention of dimensionality: A line in
2D has a point normal form, and a plane in 3D has one.
Suppose that we know line L passes through points C and B, as in Figure 4.29. What is its point normal
form? If we can find a vector n that is perpendicular to the line, then for any point R = (x, y) on the line the
vector R - C must be perpendicular to n, so we have the condition on R:
n

B
R
A
Figure 4.29. Finding the point normal form for a line.
n (R - C) = 0

(point normal form)

(4.42)

This is the point normal equation for the line, expressing that a certain dot product must turn out to be zero
for every point R on the line. It employs as data any point lying on the line, and any normal vector to the
line.

We still must find a suitable n. Let b = B - C denote the vector from C to B. Then b will serve well as the
desired n. For purposes of building the point normal form, any scalar multiple of b works just as well for
n.
Example 4.5.4. Find the point normal form. Suppose line L passes through points C = (3, 4) and B = (5, 2). Then b = B - C = (2, -6) and b = (6, 2) (sketch this). Choosing C as the point on the line, the point
normal form is: (6, 2) . ((x, y) - (3, 4)) = 0, or 6x + 2y = 26. Both sides of the equation can be divided by 26
(or any other nonzero number) if desired.
Its also easy to find the normal to a line given the equation of the line, say, f x + g y = 1. Writing this once
again as (f, g) (x, y) = 1 it is clear that the normal n is simply (f, g) (or any multiple thereof). For instance,
the line given by 5x - 2y = 7 has normal vector (5, -2), or more generally K(5, -2) for any nonzero K.
Its also straightforward to find the parametric form for a line if you are given its point normal form.
Suppose it is known that line L has point normal form n (P - C) = 0, where n and C are given explicitly.
The parametric form is then L(t) = C + nt (why?). You can also obtain the parametric form if the equation
of the line is given. a). find the normal n as in the previous paragraph, and b). find a point (Cx, Cy) on the
line by choosing any value for Cx and use the equation to find the corresponding Cy.
Moving from each representation to the others.
We have described three different ways to characterize a line. Each representation uses certain data that
distinguishes one line from another. This is the data that would be stored in a suitable data structure within
a program to capture the specifics of each line being stored. For instance, the data associated with the
representation that specifies a line parametrically as in C + bt would be the point C and the direction b. We
summarize this by saying the relevant data is {C, b}.

Hill - Chapter 4

09/23/99

page 29

The three representations and their data are:

The two point form: say C and B; data = {C, B}
The parametric form: C + bt; data = {C, b}.
The point normal (implicit) form (in 2D only): n (P - C) = 0; data = {C, n}.
Note that a point C on the line is common to all three forms. Figure 4.30 shows how the data in each
representation can be obtained from the data in the other representations. For instance, given {C, b} of the
parametric form, the normal n of the point normal form is obtained simply as b.

Figure 4.30. Moving between representations of a line.

Practice Exercise 4.5.5. Find the point normal form. Find the point normal form for the line that passes
through (-3, 4) and (6, -1). Sketch the line and its normal vector on graph paper.

Planes in 3D space.
Because there is such a heavy use of polygons in 3D graphics, planes seem to appear everywhere. A
polygon (a faceof an object) lies in its parent plane, and we often need to clip objects against planes, or
find the plane in which a certain face lies.
Planes, like lines, have three fundamental forms: the three-point form, the parametric representation and
the point normal form. We examined the three-point form in Section 4.4.2.
The parametric representation of a plane.
The parametric form for a plane is built on three ingredients: one of its points, C, and two (nonparallel)
vectors, a and b, that lie in the plane, as shown in Figure 4.31. If we are given the three (non-collinear)
points A, B, and C in the plane, then take a = A - C and b = B - C.

Hill - Chapter 4

09/23/99

page 30

b
C
a

Figure 4.31. Defining a plane parametrically.

To construct a parametric form for this plane, note that any point in the plane can be represented by a
vector sum: C plus some multiple of a plus some multiple of b. Using parameters s and t to specify the
multiples we have C + s a + t b. This provides the desired parametric form P(s, t)
P(s, t) = C + a s + b t

(4.43)

Given any values of s and t we can identify the corresponding point on the plane. For example, the position
at s = t = 0 is C itself, and that at s = 1 and t = - 2 is P(1, - 2) = C + a - 2 b.
Note that two parameters are involved in the parametric expression for a surface, whereas only one
parameter is needed for a curve. In fact if one of the parameters is fixed, say s = 3, then P(3, t) is a function
of one variable and represents a straight line: P(3, t) = (C + 3 a) + b t.
It is sometimes handy to arrange the parametric form into its component form by collecting terms
P(s, t) = (Cx + ax s + bx t, Cy + ay s + by t , Cz + az s + bz t).

(4.44)

We can rewrite the parametric form in Equation 4.43 explicitly in terms of the given points A, B, and C:
just use the definitions of a and b:
P(s, t) = C + s(A - C) + t(B - C)
which can be rearranged into the affine combination of points:
P(s, t) = s A + t B + (1 - s - t)C

(4.45)

Example 4.5.6. Find a parametric form given three points in a plane. Consider the plane passing through A
= (3,3,3), B = (5,5,7), and C = (1, 2, 4). From Equation 4.43 it has parametric form
P(s, t) = (1, 2, 4) + (2, 1, - 1) s + (4, 3, 3) t. This can be rearranged to the component form: P(s, t) = (1 + 2 s + 4
t) i + (2 + s + 3 t) j + (4 - s + 3 t) k, or to the affine combination form P(s, t) = s(3, 3, 3) + t(5, 5, 7) + (1 - s - t)(1,
2, 4).
The point normal form for a plane.
Planes can also be represented in point normal form, and the classic equation for a plane emerges at once.
Figure 4.32 shows a portion of plane P in three dimensions. A plane is completely specified by giving a
single point, B = (bx, by, bz), that lies within it, and the normal direction, n = (nx, ny, nz), to the plane. Just
as the normal vector to a line in two dimensions orients the line, the normal to a plane orients the plane in
space.

Hill - Chapter 4

09/23/99

page 31

z
n

B
x

Figure 4.32. Determining the equation of a plane.

The normal n is understood to be perpendicular to any line lying in the plane. For an arbitrary point R = (x,
y, z) in the plane, the vector from R to B must be perpendicular to n, giving:
n (R - B) = 0

(4.46)

This is the point normal equation of the plane. It is identical in form to that for the line: a dot product set
equal to 0. All points in a plane form vectors with B that have the same dot product with the normal vector.
By spelling out the dot product and using n = (nx, ny, nz), we see that the point normal form is the
traditional equation for a plane:
nx x + ny y + nz z = D

(4.47)

where D = n (B - 0). For example, if given the equation for a plane such as 5x - 2y + 8z = 2, you know
immediately that the normal to this plane is (5, -2, 8) or any multiple of this. (How do you find a point in
this plane?)
Example 4.5.7. Find a point normal form. Let plane P pass through (1, 2, 3) with normal vector (2, -1, 2).
Its point normal form is (2, -1, -2) ((x, y, z) - (1, 2, 3)) = 0. The equation for the plane may be written out
as 2x - y - 2z = 6.
Example 4.5.8. Find a parametric form given the equation of the plane. Find a parametric form for the
plane 2 x - y + 3 z = 8. Solution: By inspection the normal is (2, - 1, 3). There are many parametrizations;
we need only find one. For C, choose any point that satisfies the equation; C = (4, 0, 0) will do. Find two
(noncollinear) vectors, each having a dot product of 0 with (2, - 1, 3); some hunting finds that a = (1, 5, 1)
and b = (0, 3, 1) will work. Thus the plane has parametric form P(s, t) = (4, 0, 0) + (1, 5, 1) s + (0, 3, 1) t.
Example 4.5.9. Finding two noncollinear vectors. Given the normal n to a plane, what is an easy way to
find two noncollinear vectors a and b that are both perpendicular to n? (In the previous exercise we just
invented two that work.) Here we use the fact that the cross product of any vector with n is normal to n. So
we take a simple choice such as (0, 0, 1), and construct a as its cross product with n:
a = (0, 0, 1) n = ( -ny, nx, 0)
(Is this indeed normal to n?). We can use the same idea to form b that is normal to both n and a:
b = n a = (-nx nz, -ny nz, nx2 + ny2)
(Check that b a and b n.) So b is certainly not collinear with a.
We apply this method to the plane (3, 2, 5) (R- (2,7,0)) = 0. Set a = (0, 0, 1) n = (-2, 3, 0) and b = (-15,
-10, 13). The plane therefore has parametric form:
P(s, t) = (2 -2 s - 15 t, 7 + 3 s - 10 t, 13 t).
Check: Is P(s, t) - C = ( -2s - 15t, -3s -10 t, 13t) indeed normal to n for every s and t?
Practice Exercise 4.5.7. Find the Plane. Find a parametric form for the plane coincident with the y, zplane.

Hill - Chapter 4

09/23/99

page 32

Moving from each representation to the others.

Just as with lines, it is useful to be able to move between the three representations of a plane, to manipulate
the data that describes a plane into the form best suited to a problem.
For a plane, the three representations and their data are:
The three point form: say C , B, and A; data = {C, B, A}
The parametric form: C + as + bt; data = {C, a, b}.
The point normal (implicit) form: n (P - C) = 0; data = {C, n}.
A point C on the plane is common to all three forms. Figure 4.33 shows how the data in each representation
can be obtained from the data in the other representations. Check each one carefully. Most of these cases
have been developed explicitly in Section 4.4.2 and this section. Some are developed in the exercises. The
trickiest is probably the calculation in Example 4.5.10. Another that deserves some explanation is finding
three points in a plane when given the point normal form. One point, C, is already known. The other two
are found using special values in the point normal form itself, which is the equation nxx + nyy + nzz = n C.
Choose ,for convenience, A = (0, 0, az), and use the equation to determine that az = n C /nz. Similarly,
choose B = (0, by, 0), and use the equation to find by= n C /ny.

Figure 4.33. Moving between representations of a plane.

Planar Patches.
Just as we can restrict the parameter t in the representation of a line to obtain a ray or a segment, we can
restrict the parameters s and t in the representation of a plane.
In the parametric form of Equation 4.43 the values for s and t can range from - to , and thus the plane
can extend forever. In some situations we want to deal with only a piece of a plane, such as a
parallelogram that lies in it. Such a piece is called a planar patch, a term that invites us to imagine the
plane as a quilt of many patches joined together. Later we examine curved surfaces made up of patches
which are not necessarily planar. Much of the practice of modeling solids involves piecing together patches
of various shapes to form the skin of an object.

Hill - Chapter 4

09/23/99

page 33

A planar patch is formed by restricting the range of allowable parameter values for s and t. For instance,
one often restricts s and t to lie only between 0 and 1. The patch is positioned and oriented in space by
appropriate choices of a, b, and C. Figure 4.34a shows the available range of s and t as a square in
parameter space, and Figure 4.34b shows the patch that results from this restriction in object space.
a).
u

b).

Parameter
space

@(0, 1)

b
C

@(1, 1)

@(0, 0)

a @(1, 0)

u
world
coordinates

Figure 4.34. Mapping between two spaces to define a planar patch.

To each point (s, t) in parameter space there corresponds one 3D point in the patch P(s, t) = C + as + bt.
The patch is a parallelogram whose corners correspond to the four corners of parameter space and are
situated at
P(0, 0) = C;
P(1, 0) = C + a;
P(0, 1) = C + b;
P(1, 1) = C + a + b.

(4.48)

The vectors a and b determine both the size and the orientation of the patch. If a and b are perpendicular,
the grid will become rectangular, and if in addition a and b have the same length, the grid will become
square. Changing C just shifts the patch without changing its shape or orientation.
Example 4.5.10. Make a patch. Let C = (1, 3, 2), a = (1, 1, 0), and b = (1, 4, 2). Find the corners of the
planar patch. Solution: From the preceding table we obtain the four corners: P(0, 0) = (1, 3, 2), P(0, 1) =
(2, 7, 4), P(1, 0) = (2, 4, 2), and P(1, 1) = (3, 8, 4).
Example 4.5.11. Characterize a Patch. Find a, b, and C that create a square patch of length 4 on a side
centered at the origin and parallel to the x, z-plane. Solution: The corners of the patch are at (2, 0, 2), (2, 0,
- 2), ( - 2, 0, 2), and ( - 2, 0, - 2). Choose any corner, say (2, 0, - 2), for C. Then a and b each have length 4
and are parallel to either the x- or the z-axis. Choose a = ( - 4, 0, 0) and b = (0, 0, 4).
Practice Exercise 4.5.8. Find a Patch. Find point C and some vectors a and b that create a patch having
the four corners ( - 4, 2, 1), (1, 7, 4), ( - 2, - 2, 2), and (3, 3, 5).

4.6. Finding the Intersection of two Line Segments.

We often need to compute where two line segments in 2D space intersect. It appears in many other tasks,
such as determining whether or not a polygon is simple. Its solution will illustrate the power of parametric
forms and dot products.
The Problem: Given two line segments, determine whether they intersect, and if they do, find their point
of intersection.

Hill - Chapter 4

09/23/99

page 34

Suppose one segment has endpoints A and B and the other segment has endpoints C and D. As shown in
Figure 4.35 the two segments can be situated in many different ways: They can miss each other (a and b),
overlap in one point (c and d), or even overlap over some region (e). They may or may not be parallel. We
need an organized approach that handles all of these possibilities.
b).

a).

c).
D

D
A

A
A

C
d).

D
B

e).
A

Figure 4.35. Many cases for two line segments.

Every line segment has a parent line, the infinite line of which it is part. Unless two parent lines are
parallel they will intersect at some point. We first locate this point.
We set up parametric representations for each of the line segments in question. Call AB the segment from A
to B. Then
AB(t) = A + b t

(4.49)

where for convenience we define b = B - A. As t varies from 0 to 1 all points on the finite line segment are
visited. If t is allowed to vary from - to the entire parent line is swept out.
Similarly we call the segment from C to D by the name CD, and give it parametric representation (using a
new parameter, say, u)
CD(u) = C + d u,
where d = D - C. We use different parameters for the two lines, t for one and u for the other, in order to
describe different points on the two lines independently. (If the same parameter were used, the points on the
two lines would be locked together.)
For the parent lines to intersect, there must be specific values of t and u for which the two equations above
are equal:
A + bt = C + du
Defining c = C - A for convenience we can write this condition in terms of three known vectors and two
(unknown) parameter values:
bt = c + du

(4.50)

This provides two equations in two unknowns, similar to Equation 4.22. We solve it the same way: dot both

sides with d to eliminate the term in d, giving d bt = d c . There are two main cases: the term

d b is zero or it is not.

Case 1: The term d b is not Zero.

Hill - Chapter 4

09/23/99

page 35

Here we can solve for t obtaining:

d c

d b

(4.51)

Similarly dot both sides of Equation 4.50 with b to obtain (after using one additional property of perpdot productswhich one?):

b c
u=
d b

(4.52)

Now we know that the two parent lines intersect, and we know where. But this doesnt mean that the line
segments themselves intersect. If t lies outside the interval [0, 1], segment AB doesnt reach the other
segment, with similar statements if u lies outside of [0,1]. If both t and u lie between 0 and 1 the line
segments do intersect at some point, I. The location of I is easily found by substituting the value of t in
Equation 4.49:

d c
b
I= A+
d b

(the intersection point)

(4.53)

Example 4.6.1: Given the endpoints A = (0, 6), B = (6, 1), C = (1, 3), and D = (5, 5), find the intersection if
it exists. Solution: db = -32, so t = 7/16 and u = 13/32 which both lie between 0 and 1, and so the
segments do intersect. The intersection lies at (x, y) = (21/8, 61/16). This result may be confirmed visually
by drawing the segments on graph paper and measuring the observed intersection.

Case 2: The term d b is Zero.

In this case we know d and b are parallel (why?). The segments might still overlap, but this can happen
only if the parallel parent lines are identical. A test for this is developed in the exercises.
The exercises discuss developing a routine that performs the complete intersection test on two line
segments.
Practice Exercises.

4.6.1. When the parent lines overlap. We explore case 2 above, where the term db = 0, so the parent
lines are parallel. We must determine whether the parent lines are identical, and if so whether the segments
themselves overlap.
To test whether the parent lines are the same, see whether C lies on the parent line through A and B.
a). Show that the equation for this parent line is bx (y - Ay) - by (x - Ax) = 0.
We then substitute Cx for x and Cy for y and see whether the left-hand side is sufficiently close to zero (i.e.
its size is less than some tolerance such as 10-8). If not, the parent lines do not coincide, and no intersection
exists. If the parents lines are the same, the final test is to see whether the segments themselves overlap.
b). To do this, show how to find the two values tc and td at which this line through A and B reaches C and
D, respectively. Because the parent lines are identical, we can use just the x-component. Segment AB
begins at 0 and ends at 1, and by examining the ordering of the four values 0, 1, tc, and td, we can readily
determine the relative positions of the two lines.
c). Show that there is an overlap unless both tc and td are less than 0 or both are larger than 1. If there is an
overlap, the endpoints of the overlap can easily be found from the values of tc and td.
d). Given the endpoints A = (0, 6), B = (6, 2), C = (3, 4), and D = (9, 0), determine the nature of any
intersection.

Hill - Chapter 4

09/23/99

page 36

4.6.2. The Algorithm for determining the intersection. Write the routine segIntersect() that would
be used in the context: if(segIntersect(A, B, C, D, InterPt)) <do something>
It takes four points representing the two segments, and returns 0 if the segments do not intersect, and 1 if
they do. If they do intersect the location of the intersection is placed in interPt. It returns -1 if the parent
lines are identical.
4.6.3. Testing the Simplicity of a Polygon. Recall that a polygon P is simple if there are no edge
intersections except at the endpoints of adjacent edges. Fashion a routine int isSimple(Polygon P)
that takes a brute force approach and tests whether any pair of edges of the list of vertices of the polygon
intersect, returning 0 if so, and 1 if not so. (Polygon is some suitable class for describing a polygon.) This
is a simple algorithm but not the most efficient one. See [moret91] and [preparata85] for more elaborate
attacks that involve some sorting of edges in x and y.
4.6.4. Line Segment Intersections. For each of the following segment pairs, determine whether the
segments intersect, and if so where.
1. A = (1, 4),
B = (7, 1/2),
C = (7/2, 5/2),
D = (7, 5);
2. A = (1, 4),
B = (7, 1/2),
C = (5, 0),
D = (0, 7);
3. A = (0, 7),
B = (7, 0),
C = (8, - 1),
D = (10, - 3);

4.6.1. Application of Line Intersections: the circle through three points.

Suppose a designer wants a tool that draws the unique circle that passes through three given points. The
user specifies three points A, B, and C, on the display with the mouse as suggested in Figure 4.36a, and the
circle is drawn automatically as shown in Figure 4.36b. The unique circle that passes through three points is
called the excircle or circumscribed circle, of the triangle defined by the points. Which circle is it? We
need a routine that can calculate its center and radius.
b). What it looks like
c). How to find its center
a). Which circle?

?
C

B
B
Figure 4.36. Finding the excircle.

perp.
bisector #2

S
C

perp.
bisector #1

C
B

Figure 4.35c shows how to find it. The center S of the desired circle must be equidistant from all three
vertices, so it must lie on the perpendicular bisector of each side of triangle ABC (The perpendicular
bisector is the locus of all points that are equidistant from two given points.). Thus we can determine S if
we can compute where two of the perpendicular bisectors intersect.
We first show how to find a parametric representation of the perpendicular bisector of a line segment.
Figure 4.37 shows a segment S with endpoints A and B. Its perpendicular bisector L is the infinite line that
passes through the midpoint M of segment S, and is oriented perpendicular to it. But we know that midpoint

M is given by (A + B)/2, and the direction of the normal is given by ( B A ) , so the perpendicular
bisector has parametric form:

Hill - Chapter 4

09/23/99

page 37

M
A

Figure 4.37. The perpendicular bisector of a segment.

L(t) =

1
(A + B) + (B A) t
2

(the perpendicular bisector of AB)

(4.54)

Now we are in a position to compute the excircle of three points. Returning to Figure 4.35b we seek the
intersection S of the perpendicular bisectors of AB and AC. For convenience we define the vectors:
a=B-A
b=C-B
c=A-C

(4.55)

To find the perpendicular bisector of AB we need the midpoint of AB and a direction perpendicular to AB.
The midpoint of AB is A + a / 2 (why?). The direction perpendicular to AB is a . So the parametric form
for the perpendicular bisector is A + a / 2 + a t. Similarly the perpendicular bisector of AC is A - c / 2 +

c u, using parameter u. Point S lies where these meet, at the solution of:
a t = b / 2 + c u

(where we have used a + b + c = 0). To eliminate the term in u take the dot product of both sides with c,
and obtain t = 1/2 (b c) / (a c). To find S use this value for t in the representation of the perpendicular
bisector: A + a / 2 + a t, which yields the simple explicit form10:

b c
1
S = A + a + a
a c
2

(center of the excircle)

(4.56)

The radius of the excircle is the distance from S to any of the three vertices, so it is |S - A|. Just form the
magnitude of the last term in Equation 4.56. After some manipulation (check this out) we obtain:

a
radius =
2

b c
1
a c +
2

(radius of the excircle)

(4.57)

Once S and the radius are known, we can use drawCircle() from Chapter 3 to draw the desired circle.
Example 4.6.2. Find the perpendicular bisector L of the segment S having endpoints A = (3, 5) and B = (9,
3).

Solution: By direct calculation, midpoint M = (6, 4), and (B A) = (2, 6), so L has representation L(t) =
(6 + 2t, 4 + 6t). It is useful to plot both S and L to see this result.

10Other closed form expressions for S have appeared previously, e.g. in [goldman90] and [lopex92]

Hill - Chapter 4

09/23/99

page 38

Every triangle also has an inscribed circle, which is sometimes necessary to compute in a computer-aided
design context. A case study examines how to do this, and also discusses the beguiling nine-point circle.
Practice Exercise 4.6.5. A Perpendicular Bisector. Find a parametric expression for the perpendicular
bisector of the segment with endpoints A = (0, 6) and B = (4, 0). Plot the segment and the line.

4.7. Intersections of Lines with Planes, and Clipping.

The task of finding the intersection of a line with another line or with a plane arises in a surprising variety
of situations in graphics. We have already seen one approach in Section 4.6, that finds where two line
segments intersect. That approach used parametric representations for both the line segments, and solved
two simultaneous equations.
Here we develop an alternative method that works for both lines and planes. It represents the intersecting
line by a parametric representation, and the line or plane being intersected in a point normal form. It is very
direct and clearly reveals what is going on. We develop the method once, and then apply the results to the
problem of clipping a line against a convex polygon in 2D, or a convex polyhedron in 3D. In Chapter 7 we
see that this is an essential step in viewing 3D objects. In Chapter 14 we use the same intersection
technique to get started in ray tracing.
In 2D we want to find where a line intersects another line; in 3D we want to find where a line intersects a
plane. Both of these problems can be solved at once because the formulation is in terms of dot products,
and the same expressions arise whether the involved vectors are 2D or 3D. (We also address the problem of
finding the intersection of two planes in the exercises: it too is based on dot products.)
Consider a line described parametrically as R(t) = A + c t. We also refer to it as a ray. We want to
compute where it intersects the object characterized by the point normal form n (P - B) = 0. In 2D this is a
line; in 3D it is a plane. Point B lies on it, and vector n is normal to it. Figure 4.38a shows the ray hitting a
line, and part b) shows it hitting a plane. We want to find the location of the hit point.
n
n
a).
b).
B
'hit' point
c
c
A
A
'hit' point
B
n . (P - B) = 0
n . (P - B) = 0
Figure 4.38. Where does a ray hit a line or a plane?
Suppose it hits at t = thit, the hit time. At this value of t the line and ray must have the same coordinates,
so A + c thit must satisfy the equation of the point normal form of the line or plane. Therefore we substitute
this unknown hit point into the point normal equation to obtain a condition on thit:
n (A + c thit - B) = 0.
This may be rewritten as
n (A - B) + n c thit = 0,
which is a linear equation in thit. Its solution is:

t hit =

n (B A)
n c

Hill - Chapter 4

(hit time 2D and 3D cases)

09/23/99

(4.58)

page 39

As always with a ratio of terms we must examine the eventuality that the denominator of thit is zero. This
occurs when nc = 0, or when the ray is aimed parallel to the plane, in which case there is no hit at all.11
When the hit time has been computed, it is simple to find the location of the hit point : Substitute thit into
the representation of the ray:
hit point: Phit = A + cthit

(hit spot 2D and 3D cases)

(4.59)

In the intersection problems treated below we will also need to know generally which direction the ray
strikes the line or plane: along with the normal n or counter to n. (This will be important because we
will need to know whether the ray is exiting from an object or entering it.) Figure 4.39 shows the two
possibilities for a ray hitting a line. In part a) the angle between the rays direction, c, and n is less than 900
so we say the ray is aimed along with n. In part b) the angle is greater than 90o so the ray is aimed
counter to n.
a). ray is aimed "along with" n
b). ray is aimed "counter to" n

n
c

c
A

Figure 4.39. The direction of the ray is along or against n.

It is easy to test which of these possibilities occurs, since the sign of n c tells immediately whether the
angle between n and c is less than or greater than 90o. Putting these ideas together, we have the three
possibilities:

if n c > 0
if n c = 0
if n c < 0

the ray is aimed along with the normal;

the ray is parallel to the line
the ray is aimed counter to the normal

(4.60)

Practice Exercises.
4.7.1. Intersections of rays with lines and planes. Find when and where the ray A + ct hits the object n
(P - B) = 0 (lines in the 2D or planes in the 3D).
a). A = (2, 3), c = (4, -4), n = (6,8), B = (7,7).
b). A = (2, -4, 3), c = (4, 0, -4), n = (6,9, 9), B = (-7, 2, 7).
c). A = (2, 0), c = (0 -4), n = (0,8), B = (7,0).
d). A = (2, 4, 3), c = (4, 4, -4), n = (6,4, 8), B = (7, 4, 7).
4.7.2. Rays hitting Planes. Find the point where the ray (1,5,2) + (5, -2, 6)t hits the plane 2x -4y + z = 8.
4.7.3. What is the intersection of two planes? Geometrically we know that two planes intersect in a
straight line. But which line? Suppose the two planes are given by n ( P A) = 0 and m ( P B) = 0 .
Find the parametric form of the line in which they intersect. You may find it easiest to:
a). First obtain a parametric form for one of the planes: say, C + as + bt for the second plane.
b). Then substitute this form into the point normal form for the first plane, thereby obtaining a linear
equation that relates parameters s and t.
c). Solve for s in terms of t, say s = E + Ft. (Find expressions for E and F.)
d). Write the desired line as C + a(E + Ft) + bt.

4.8. Polygon Intersection Problems.

11If the numerator is also 0 the ray lies entirely in the line (2D) or plane (3D). (why?).

Hill - Chapter 4

09/23/99

page 40

We know polygons are the fundamental objects used in both 2D and 3D graphics. In 2D graphics their
straight edges make it easy to describe them and draw them. In 3D graphics, an object is often modeled as
a polygonal mesh: a collection of polygons that fit together to make up its skin. If the skin forms a
closed surface that encloses some space the mesh is called a polyhedron. We study meshes and polyhedra
in depth in Chapter 6.
Figure 4.40 shows a 2D polygon and a 3D polyhedron that we might need to analyze or render in a graphics
application. Three important questions that arise are:

Figure 4.40. Intersection problems of a line and a polygonal object.

a). Is a given point P inside or outside the object?
b). Where does a given ray R first intersect the object?
c). Which part of a given line L lies inside the object, and which part lies outside?
As a simple example, which part(s) of the line 3y - 2x = 6 lie inside the polygon whose vertices are (0, 3), (2, -2), (-5, 0), (0, -7), (1, 1)?

4.8.1. Working with convex polygons and polyhedra.

The general case of intersecting a line with any polygon or polyhedron is quite complex; we address it in
Section 4.8.4. Things are much simpler when the polygon or polyhedron is convex. They are simpler
because a convex polygon is completely described by a set of bounding lines; in 3D a convex polyhedron
is completely described by a set of bounding planes. So we need only test the line against a set of
unbounded lines or planes.
Figure 4.41 illustrates this for the 2D case. Part a shows a convex pentagon, and part b) shows the bounding
lines L0, L1, etc. of the pentagon. Each bounding line defines two half spaces: the inside half space that
contains the polygon, and the outside half space that shares no points with the polygon. Part c) of the figure
shows a portion of the outside half space associated with the bounding line L2.

Hill - Chapter 4

09/23/99

page 41

a).

b).

c).

outside
half space

L2
L1
L0

Figure 4.41. Convex polygons and polyhedra.

Example 4.8.1. Finding the bounding lines. Figure 4.42a shows a unit square. There are four bounding
lines, given by x = 1, x = -1, y = 1, and y = -1. In addition, for each bounding line we can identify the
outward normal vector: the one that points into the outside half space of the bounding line. The outward
normal vector for the line y = 1 is of course n = (0, 1). (What are the other three?)
a).
b).
y

y
y=1
1

1
1

an outward normal

Figure 4.42. Examples of convex polygons.

The triangle in part b) has three bounding lines. (What is the equation for each line?) The point normal
form for each of the three lines is given next; in each case it uses the outward normal (check this):
(-1, 0) (P - (0, 0)) = 0;
(0 , -1) (P - (0, 0)) = 0;
(1 , 1) (P - (1, 0)) = 0;
The big advantage in dealing with convex polygons is that we perform intersection tests only on infinite
lines, and dont need to check whether an intersection lies beyond an endpoint recall the complexity
of the intersection tests in Section 4.7. In addition the point normal form can be used, which simplifies the
calculations.
For a convex polyhedron in 3D, each plane has an inside and an outside half space, and an outward
pointing normal vector. The polyhedron is the intersection of all the inside half spaces, (the set of all points
that are simultaneously in the inside half space of every bounding plane).

4.8.2. Ray Intersections and Clipping for Convex Polygons.

We developed a method in Section 4.7 that finds where a ray hits an individual line or plane. We can use
this method to find where a ray hits a convex polygon or polyhedron.
The Intersection Problem. Where does the ray A + ct hit polygon P?
Figure 4.43 shows a ray A + c t intersecting polygon P. We want to know all of the places where the ray
hits P. Because P is convex the ray hits P exactly twice: It enters once and exits once. Call the values of t at
which it enters and exits tin and tout, respectively. The ray intersection problem is to compute the values of
tin and tout. Once these hit times are known we of course know the hit points themselves:

Hill - Chapter 4

09/23/99

page 42

P0
P1
c
A

P2
n1
L1

Figure 4.43. Ray A + ct intersecting a convex polygon.

Entering hit point: A + c tin
Exiting hit point: A + c tout

(4.61)

The ray is inside P for all t in the interval [tin, tout].

Note that finding tin and tout not only solves the intersection problem, but also the clipping problem. If we
know tin and tout we know which part of the line A + ct lies inside P. Usually the clipping problem is stated
as:
The Clipping problem: For the two points A and C which part of segment AC lies inside P?
Figure 4.44 shows several possible situations. Part a) shows the case where A and C both lie outside P, but
there is a portion of the segment AC that lies inside P.

a).

b).
0

tin

tout

c).

1
1

tin

C
A

Figure 4.44. A segment clipped by a polygon.

If we consider segment AC as part of a ray given by A + ct where c = C - A, then point A corresponds to
the point on the ray at t = 0, and C corresponds to the point at t = 1. These ray times are labeled in the
figure. To find the clipped segment we compute tin and tout as described above. The segment that survives
clipping has end points A + c tin and A + c tout. In Figure 4.43b point C lies inside P and so tout is larger than
1. The clipped segment has end points A + c tin and C. In part c) both A and C lie inside P, so the clipped
segment is the same: AC.
In general we compute tin, and compare it to 0. The larger of the values 0 and tin is used as the time for
the first end point of the clipped segment. Similarly, the smaller of the values 1 and tout is used to find the
second end point. So the end points of the clipped segment are:
A = A + c max(0, tin)
C = A + c min(tout,1)

(4.62)

Now how are tin and tout computed? We must consider each of the bounding lines of P in turn, and find
where the ray A + c t intersects it. We suppose each bounding line is stored in point normal form as the pair

Hill - Chapter 4

09/23/99

page 43

{B, n}, where B is some point on the line and n is the outward pointing normal for the line: it points to the
outside of the polygon. Because it is outward pointing the test of Equation 4.60 translates to:
if n c > 0
if n c = 0
if n c < 0

the ray is exiting from P;

the ray is parallel to the line
the ray is entering P

(4.63)

For each bounding line we find:

a). The hit time of the ray with the bounding line (use Equation 4.58);
b). Whether the ray is entering or exiting the polygon (use Equation 4.63)
If the ray is entering, we know that the time at which the ray ultimately enters P (if it enters it at all) cannot
be earlier than this newly found hit time. We keep track of the earliest possible entering time as tin. For
each entering hit time, thit, we replace tin by max(tin, thit). Similarly we keep track of the latest possible exit
time as tout, and for each exiting hit we replace tout by min(tout, thit).
It helps to think of the interval [tin, tout] as the candidate interval of t, the interval of t inside of which the
ray might lie inside the object. Figure 4.45 shows an example for the clipping problem. We know the point
A + ct cannot be inside P for any t in the candidate interval. As each bounding line is tested, the candidate
interval gets reduced as tin is increased or tout is decreased: pieces of it get chopped off. To get started we
initialize tin to 0 and tout to 1 for the line clipping problem, so the candidate interval is [0,1].

candidate
interval

the ray is
outside P here

the ray is
outside P here
t

tin

tout

Figure 4.45. The candidate interval for a hit.

The algorithm is then:
1). Initialize the candidate interval to [0,1]12.
2). For each bounding line, use Equation 4.58 to find the hit time thit and determine whether its an entering
or exiting hit:
if its an entering hit, set tin = max(tin, thit)
if its an exiting hit, set tout = min(tout thit)
If at any point tin becomes greater than tout we know the ray misses P entirely, and testing is terminated
3). If candidate interval is not empty, then from Equation 4.62 the segment from A + c tin to A + c tout is
known to lie inside P. For the line clipping problem these are the endpoints of the clipped line. For the ray
intersection problem we know the entering and exiting points of the ray.
Note that we stop further testing as soon the candidate interval vanishes. This is called an early out: if we
determine early in the processing that the ray is outside of the polygon, we save time by immediately
exiting from the test.
Figure 4.46 shows a specific example of clipping: we seek the portion of segment AC that lies in polygon
P. We initialize tin to 0 and tout to 1. The ray starts at A at t = 0 and proceeds to point C, reaching it at t =
1. We test it against each bounding line L0, L1, .., in turn and update tin and tout as necessary.

For the ray intersection problem, where the ray extends infinitely far in both directions, we set tin = -
and tout = . In practice tin is set to a large negative value, and tout to a large positive value.
12

Hill - Chapter 4

09/23/99

page 44

@.2
@0
L5
L0

A
L4

@.28
@1

intersects L3
@ -4.7

@.66

C
@.83

intersects L2
@3.4

Figure 4.46. Testing when a ray lies inside a convex polygon.

Suppose when we test it against line L0 we find an exiting hit at t = 0 .83. This sets tout to 0.83, and the
candidate interval is now [0, 0.83]. We then test it against L1 and find an exiting hit at t = 0.66. This reduces
the candidate interval to [0, 0.66]. The test against L2 gives an exiting hit at t = 3.4. This tells us nothing
new: we already know the ray is outside for t > 0.66. The test against L3 gives an entering hit at t = -0.47.
So we set tin to -0.47, and the candidate interval is [-0.47, 0.66]. The test with L4 gives an entering hit at t =
0.2, so tin is updated to 0.2. Finally, testing against L5 gives an entering hit at t = 0.28, and we are done. The
candidate interval is [0.28, 0.66].In fact the ray is inside P for all t between 0.28 and 0.66.
Figure 4.47 shows the sequence of updates to tin and tout that occur as each of the lines above is tested.
line test
tin
tout
0
0
0.83
1
0
0.66
2
0
0.66
3
0
0.66
4
0.2 0.66
5
0.28 0.66
Figure 4.47. Updates on the values of tin and tout.

4.8.3. The Cyrus-Beck Clipping Algorithm.

We build a routine from these ideas, that performs the clipping of a line segment against any convex
polygon. The method was originally developed by Cyrus and Beck [cyrus78]. Later a highly efficient
clipper for rectangular windows was devised by Liang and Barsky [liang84] based on similar ideas. It is
discussed in a Case Study at the end of this chapter.
The routine that implements the Cyrus-Beck clipper has interface:
int CyrusBeckClip(Line& seg, LineList& L);
Its parameters are the line segment, seg, to be clipped (which contains the first and second endpoints
named seg.first and seg.second) and the list of bounding lines of the polygon. It clips seg against
each line in L as described above, and places the clipped segment back in seg. (This is why seg must be
passed by reference.) The routine returns:

0 if no part of the segment lies in P ( the candidate interval became empty);

1 if some part of the segment does lie in P.

Figure 4.48 shows pseudocode for the Cyrus Beck algorithm. The types LineSegment, LineList, and
Vector2 are suitable data types to hold the quantities in question (see the exercises). Variables numer
and denom hold the numerator and denominator for thit of Equation 4.48:

Hill - Chapter 4

09/23/99

page 45

numer = n ( B A)
denom = n c

(4.64)

int CyrusBeckClip(LineSegment& seg, LineList L)

{
double numer, denom; // used to find hit time for each line
double tIn = 0.0, tOut = 1.0;
Vector2 c, tmp;
form vector: c = seg.second - seg.first
for(int i = 0; i < L.num; i++) // chop at each bounding line
{
form vector tmp = L.line[i].pt - first
numer = dot(L.line[i].norm, tmp);
denom = dot(L.line[i].norm, c);
if(!chopCI(numer, denom, tIn, tOut)) return 0; // early out
}
// adjust the endpoints of the segment; do second one 1st.
if (tOut < 1.0 ) // second endpoint was altered
{
seg.second.x = seg.first.x + c.x * tOut;
seg.second.y = seg.first.y + c.y * tOut;
}
if (tIn > 0.0) // first endpoint was altered
{
seg.first.x = seg.first.x + c.x * tIn;
seg.first.y = seg.first.y + c.y * tIn;
}
return 1; // some segment survives
}
Figure 4.48. Cyrus-Beck Clipper for a Convex Polygon, 2D case (pseudocode).
Note that the value of seg.second is updated first, since we must use the old value of seg.first in
the update calculation for both seg.first and seg.second.
The routine chopCI() is shown in Figure 4.49. It uses numer and denom of Equation 4.64 to calculate
the hit time at which the ray hits a bounding line, uses Equation 4.63 to determine whether the ray is
entering or exiting the polygon, and chops off the piece of the candidate interval CI that is thereby found
to be outside the polygon.
int chopCI(double& tIn, double& tOut, double numer, double
denom)
{
double tHit;
if (denom < 0)
// ray is entering
{
tHit = numer / denom;
if (tHit > tOut) return 0; // early out
else if (tHit > tIn) tIn = tHit; // take larger t
}
else if(denom > 0)
// ray is exiting
{
tHit = numer / denom;
if(tHit < tIn) return 0;
// early out
if(tHit < tout) tOut = tHit; // take smaller t
}
else
// denom is 0: ray is parallel
if(numer <= 0) return 0;
// missed the line
return 1; // CI is still non-empty
}
Figure 4.49. Clipping against a single bounding line.

Hill - Chapter 4

09/23/99

page 46

If the ray is parallel to the line it could lie entirely in the inside half space of the line, or entirely out of it. It
turns out that numer = n (B - A) is exactly the quantity needed to tell which of these cases occurs. See the
exercises.
The 3D case: Clipping a line against a Convex Polyhedron.
The Cyrus Beck clipping algorithm works in three dimensions in exactly the same way. In 3D the edges of
the window become planes defining a convex region in three dimensions, and the line segment is a line
suspended in space. ChopCI() needs no changes at all (since it uses only the values of dot products through numer and denom ). The data types in CyrusBeckClip() must of course be extended to 3D
types, and when the endpoints of the line are adjusted the z-component must be adjusted as well.
Practice Exercises.
4.8.2. Data types for variable in the Cyrus Beck Clipper. Provide useful definitions for data types, either as
structs or classes, for LineSegment, LineList, and Vector2 used in the Cyrus Beck clipping
algorithm.
4.8.3. What does numer <= 0 do?
Sketch the vectors involved in value of numer in chopCI() and show that when the ray A + c t moves
parallel to the bounding line n (P - B) = 0, it lies wholly in the inside half space of the line if and only if
numer > 0.
4.8.4. Find the Clipped Line. Find the portion of the segment with endpoints (2, 4) and (20, 8) that lies within
the quadrilateral window with corners at (0, 7), (9, 9), (14,4), and (2, 2).

4.8.4. Clipping against arbitrary polygons.

We saw how to clip a line segment against a convex polygon in the previous section. We generalize this to
a method for clipping a segment against any polygon.
The basic problem is to find where the ray A + ct lies inside polygon P given by the vertex list P0, P1, .., PNFigure 4.50 shows an example.

P4
P3

Figure 4.50. Where is a ray inside an arbitrary polygon P?

It is clear that the ray can enter and exit from P multiple times in general, and that the result of clipping a
segment against P may result in a list of segments rather than a single one. Also, of course, P is no longer
described by a collection of infinite bounding lines in point normal form; we must work with the N finite
segments such as P3 P4 that form its edges.
The problem is close to the problem we dealt with in Section 4.7: finding the intersection of two line
segments. Now we are intersecting one line segment with the sequence of line segments associated with P.
We represent each edge of P parametrically (rather than in point normal form). For instance, the edge P3 P4
is represented as P3 + e3u where e3 = P4 - P3 is the edge vector associated with P3. In general, the i-th edge
is given by Pi + eiu, for u in [0,1] and i = 0, 1, ..., N-1where ei = Pi+1 - Pi , and as always we equate PN with
P0.
Recall from Section 4.7 that the ray A + ct hits the i-th edge when t and u have the proper values to make A
+ ct = Pi + ei u. Calling vector bi = Pi - A we seek the solution (values of t and u) of

Hill - Chapter 4

09/23/99

page 47

ct = bi + eiu
Equations 4.51 and 4.52 hold the answers. When converted to the current notation we have:

ei bi
t=
ei c

and

c bi
u= .
ei c

If e i c is 0 the i-th edge is parallel to the ray direction c and there is no intersection. There is a true
intersection with the i-th edge only if u falls in the interval [0,1].
We need to find all of the legitimate hits of the ray with edges of P, and place them in a list of the hit times.
Call this list hitList. Then pseudocode for the process would look like:

initialize hitList to empty

for(int i = 0; i < N; i++)
// for each edge of P
{
build bi, ei for the i-th edge
solve for t, u
if(u lies in [0,1])
add t to the hitList
}
What we do now with this list depends on the problem at hand.
The ray intersection problem. (Where does the ray first hit P?)
This is solved by finding the smallest value of t, tmin, in theList. The hit spot is, as always, A + c tmin.
The line clipping problem.
For this we need the sequence of t-intervals in which the ray is inside P. This requires sorting theList
and then taking the t-values in pairs. The ray enters P at the first time in each pair, and exits from P at the
second time of each pair.
Example 4.9.2. Clip AB to polygon P. Suppose the line to be clipped is AB as shown in Figure 4.51, for
which A = (1, 1) and B = (8, 2).

(6,2)

P0 = (3,2)

(8,2)
B

A
(4,1)
1
P1
(6,-1)

Figure 4.51. Clipping a line against a polygon.

P is given by the vertex list: (3, 2), (2, 0), (6, -1), (6, 2), (4, 1). Taking each edge in turn we get for the
values of t and u at the intersections:
edge
0
1
2
3

u
0.3846
-0.727
0.9048
0.4

Hill - Chapter 4

R
Lb

B
Lb

Figure 4.53. The inscribed circle of ABC is the excircle of RST.

As was the case with the excircle, the hard part is finding the center of the inscribed circle. A
straightforward method14 recognizes that the inscribed circle of ABC is simply the excircle of a different set
of three points, RST as shown in Figure 4.53a.
We need only find the locations of R, S, and T and then use the excircle method of Section 4.6.1. Figure
4.53b shows the distances of R, S, and T from A, B, and C. By the symmetry of a circle the distances |B - R|
and |B - S| must be equal, and there are two other pairs of lines that have the same length. We therefore
have (using the definitions of Equation 4.55 for a, b, and c):
|a| = Lb + La,

|b| = Lb + Lc,

|c| = La + Lc

which can be combined to solve for La and Lb:

2 La = |a| + |c| - |b|,

2 Lb = |a| + |b| - |c|

so La and Lb are now known. Thus R, S, and T are given by:

13Note: finding the incircle also solves the problem of finding the unique circle that is tangent to 3 noncollinear lines in

the plane.
14 Suggested by Russell Swan.

Hill - Chapter 4

09/23/99

page 51

a
| a|
b
S = B + Lb
|b|
c
T = A La
|c|
R = A + La

(4.65)

(Check these expressions!)

Encapsulate the calculation of R, S, and T from A, B, and C in a simple routine having usage
getTangentPoints(A, B, C, R, S, T).The advantage here is that if we have a routine
excircle() that takes three points and computes the center and radius of the excircle defined by them,
we can use the same routine to find the inscribed circle. Experiment with these tools.
The nine-point circle.
For any triangle, there are nine particularly distinguished points:
the midpoints of the 3 sides;
the feet of the 3 altitudes;
the midpoints of the lines joining the orthocenter (where the 3 altitudes meet) to the vertices.
Remarkably, a single circle passes through all nine points! Figure 4.54 shows the 9-point circle15 for an
example triangle. The nine-point circle is perhaps most easily drawn as the excircle of the midpoints of the
sides of the triangle.

Figure 4.54. The 9-point circle.

4.10.3. Case Study 4.3. Is point Q inside convex polygon P?

(Level of Effort: II.) Suppose you are given the specification of a convex polygon, P. Then given a point Q
you are asked to determine whether or not Q lies inside P. But from the discussion on convex polygons in
Section 4.8.1 we know this is equivalent to asking whether Q lies on the inside half space of every
bounding line of P. For each bounding line Li we need only test whether the vector Q - Pi is more than 900
away from the outward pointing normal.
Fact: Q lies in P if (Q - Pi) ni < 0

for i = 0, 1, ..., N-1.

(4.66)

Figure 4.55 illustrates the test for the particular bounding line that passes through P1 and P2. For the case of
point Q, which lies inside P, the angle with n1 is greater than 900. For the case of point Q which lies
outside P the angle is less than 900.

15This circle is the first really exciting one to appear in any couse on elementary geometry. Daniel Pedoe. Circles,
Pergamon Press, New York, 1957

Hill - Chapter 4

09/23/99

page 52

n0
n1

Q'
P2
Q
Figure 4.55. Is point Q inside polygon P?
Write and test a program that allows the user to:
a). lay down the vertices of a convex polygon, P, with the mouse;
b). successively lay down test points, Q, with the mouse;
c). prints is inside or is not inside depending on whether the point Q is or is not inside P.

4.10.4. Case Study 4.4. Reflections in a Chamber (2D Ray Tracing)

(Level of Effort: II.) This case study applies some of the tools and ideas introduced in this chapter to a
fascinating yet simple simulation. The simulation performs a kind of ray tracing, based in a 2D world for
easy visualization. Three dimensional ray tracing is discussed in detail in Chapter 14.
This simulation traces the path of a single tiny pinball as it bounces off various walls inside a chamber.
Figure 4.56a shows a cross section of a convex chamber W that has six walls and contains three convex
pillars. The pinball begins at point S and moves in a straight line in direction c until it hits a barrier,
whereupon it reflects off the barrier and moves in a new direction, again in a straight line. It continues to
do this forever. Figure 4.56b shows an example of the polyline path that a ray traverses.

a).

b).
c

i
p

c).

inside

outside

d).

inside

outside

i
p

Figure 4.59. Four cases for each edge of S.

a. Both s and p are inside: p is output.
b. s is inside and p is outside. Find the intersection i and output it.
c. Both s and p are outside. Nothing is output.
d. s is outside and p is inside. Find intersection i, and output i and then p.
Now follow the progress of the SutherlandHodgman algorithm in Figure 4.58. Consider clipping S against
the top edge of C. The input vertex list for this phase is a b c d e f g. The first edge from the list is taken for
convenience as that from g to a, the edge that wraps around from the end of the list to its first element.
Thus point s is g and point p is a here. Edge g, a, meaning the edge from g to a, intersects the clip edge at a
new point 1, which is output to the new list. (The output list from each stage in the algorithm is shown
below the subsequent figure in Figure 4.58.) The next edge in the input list is a, b. Since both endpoints are
above the clipping edge, nothing is output. The third edge, b, c, generates two output points, 2 and c, and
the fourth edge, c, d, outputs point d. This process continues until the last edge, f, g, is tested, producing g.
The new vertex list for the next clipping stage is therefore 1 2 c d e f g. It is illuminating to follow the
example in Figure 4.58 carefully in its entirety to see how the algorithm works.
Notice that extraneous edges 3, 6 and 9, 10 are formed that connect the three polygon fragments formed in
the clipping algorithm. Such edges can cause problems in some polygon filling algorithms. It is possible
but not trivial to remove these offending edges [sutherland74].
Task: Implement the Sutherland-Hodgman clipping algorithm, and test it on a variety of sample polygons.
The user lays down the convex polygon C with the mouse, then lays down the subject polygon S with the
mouse. It is drawn in red as it is being laid down. Clipping is then performed, and the clipped polygon(s)
are drawn in blue.

4.10.7. Case Study 4.7. Clipping a Polygon against another Weiler Atherton
Clipping.
(Level of Effort: III). This method provides the most general clipping mechanism of all we have studied. It
clips any subject polygon against any (possibly non-convex) clip polygon. The polygons may even contain
holes.

Hill - Chapter 4

09/23/99

page 56

The Sutherland-Hodgman algorithm examined in Case Study 4.6 exploits the convexity of the clipping
polygon through the use of inside-outside half-spaces. In some applications, such as hidden surface
removal and rendering shadows, however, one must clip one concave polygon against another. Clipping is
more complex in such cases. The WeilerAtherton approach clips any polygon against any other, even
when they have holes. It also allows one to form the set theoretic union, intersection, and difference of
two polygons, as we discuss in Case Study 4.8.
We start with a simple example, shown in Figure 4.60. Here two concave polygons, SUBJ and CLIP, are
represented by the vertex lists, (a, b, c, d) and (A, B, C, D), respectively. We adopt the convention here of
listing vertices so that the interior of the polygon is to the right of each edge as we move cyclically from
vertex to vertex through the list. For instance, the interior of SUBJ lies to the right of the edge from c to d
and to the right of that from d to a. This is akin to listing vertices in clockwise order.
a
SUBJ
C
A
6

5
3
4

1
2

d
b

CLIP

D
Figure 4.60 .WeilerAtherton clipping.

All of the intersections of the two polygons are identified and stored in a list (see later). For the example
here, there are six such intersections. Now to clip SUBJ against CLIP, traverse around SUBJ in the
forward direction (i.e., so that its interior is to the right) until an entering intersection is found: one for
which SUBJ is moving from the outside to the inside of CLIP. Here we first find 1, and it goes to an output
list that records the clipped polygon(s).
The process is now simple to state in geometric terms: Traverse along SUBJ, moving segment by segment,
until an intersection is encountered (2 in the example). The idea now is to turn away from following SUBJ
and to follow CLIP instead. There are two ways to turn. Turn so that CLIP is traversed in its forward
direction. This keeps the inside of both SUBJ and CLIP to the right. Upon finding an intersection, turn and
follow along SUBJ in its forward direction, and so on. Each vertex or intersection encountered is put on the
output list. Repeat the turn and jump between polygons process, traversing each polygon in its forward
direction, until the first vertex is revisited. The output list at this point consists of (1, b, 2, B).
Now check for any other entering intersections of SUBJ. Number 3 is found and the process repeats,
generating output list (3, 4, 5, 6). Further checks for entering intersections show that they have all been
visited, so the clipping process terminates, yielding the two polygons (1, b, 2, B) and (3, 4, 5, 6). An
organized way to implement this follow in the forward direction and jump process is to build the two lists
SUBJLIST: a, 1, b, 2, c, 3, 4, d, 5, 6
CLIPLIST: A, 6, 3, 2, B, 1, C, D, 4, 5
that traverse each polygon (so that its interior is to the right) and list both vertices and intersections in the
order they are encountered. (What should be done if no intersections are detected between the two

Hill - Chapter 4

09/23/99

page 57

polygons?) Therefore traversing a polygon amounts to traversing a list, and jumping between polygons is
effected by jumping between lists.
Notice that once the lists are available, there is very little geometry in the processjust a point outside
polygon test to properly identify an entering vertex. The proper direction in which to traverse each
polygon is embedded in the ordering of its list. For the preceding example, the progress of the algorithm is
traced in Figure 4.61.
start
restart

SUB_LIST:

CLIP_LIST:
visited

visited

= output the point

Figure 4.61. Applying the WeilerAtherton method.

A more complex example involving polygons with holes is shown in Figure 4.62. The

fsh: old Figure A6.6

Figure 4.62. WeilerAtherton clipping: polygons with holes.

vertices that describe holes are also listed in order such that the interior of the polygon lies to the right of an
edge. (For holes this is sometimes called counterclockwise order.) The same rule is used as earlier: Turn
and follow the other polygon in its forward direction. Beginning with entering intersection 1, the polygon
(1, 2, 3, 4, 5, i, 6, H) is formed. Then, starting with entering intersection 7, the polygon (7, 8, 9, c, 10, F) is
created. What entering intersection should be used to generate the third polygon? It is a valuable exercise to
build SUBJLIST and CLIPLIST and to trace through the operation of the method for this example.
As with many algorithms that base decisions on intersections, we must examine the preceding method for
cases where edges of CLIP and SUBJ are parallel and overlap over a finite segment.
Task: Implement the Weiler-Atherton clipping algorithm, and test it on a variety of polygons. Generate
SUBJ and CLIP polygons, either in files or by letting the user lay down polygons with the mouse. In your
implementation carefully consider how the algorithm will operate in situations such as the following:
Some edges of SUBJ and CLIP are parallel and overlap over a finite segment,
SUBJ or CLIP or both are nonsimple polygons,

Hill - Chapter 4

09/23/99

page 58

Some edges of SUBJ and CLIP overlap only at their endpoints,

CLIP and SUBJ are disjoint,
SUBJ lies entirely within a hole of CLIP.

4.10.8. Case Study 4.8. Boolean Operations on Polygons.

(Level of Effort: III.) If we view polygons as sets of points (the set of all points on the boundary or in the
interior of the polygon), then the result of the previous clipping operation is the intersection of the two
polygons, the set of all points that are in both CLIP and SUBJ. The polygons output by the algorithm
consist of points that lie both within the original SUBJ and within the CLIP polygons. Here we generalize
from intersections to other set theoretic operations on polygons, often called Boolean operations. Such
operations arise frequently in modeling [mortenson85] as well as in graphics (see Chapter 14). In general,
for any two sets of points A and B, the three set theoretic operations are
intersection: A B = {all points in both A and B}
union: A B = {all points in A or in B or both}
difference: A - B= {all points in A but not B}
with a similar definition for the set difference B - A. Examples of these sets are shown in Figure 4.63.
AXB
A+B
A-B
B-A

A
B

Figure 4.63. Polygons formed by boolean operations on polygons.

It is not hard to adjust the Weiler-Atherton method, which already performs intersections, to perform the
union and difference operations on polygons A and B.
1. Computing the union of A and B. Traverse around A in the forward direction until an exiting intersection
is found: one for which A is moving from the inside to the outside of B. Output the intersection and traverse
along A until another intersection with B is found. Now turn to follow B in its forward direction. At each
subsequent intersection, output the vertex and turn to follow the other polygon in its forward direction.
Upon returning to the initial vertex, look for other exiting intersections that have not yet been visited.
2. Computing the difference A - B(outside clipping). Whereas finding the intersection of two polygons
results in clipping one against the other, the difference operation shields one polygon from another. That
is, the difference SUBJ - CLIP consists of the parts of SUBJ that lie outside CLIP. No parts of SUBJ are
drawn that lie within the border of CLIP, so the region defined by CLIP is effectively protected, or
shielded.
Traverse around A until an entering intersection into B is found. Turn to B, following it in the reverse
direction, (so that B's interior is to the left). Upon reaching another intersection, jump to A again. At each
intersection, jump to the other polygon, always traversing A in the forward direction and B in the reverse
direction. Some examples of forming the union and difference of two polygons are shown in Figure 4.64.
The three set operations generate the following polygons:
POLYA POLYB:
4, 5, g, h(a hole)
8, B, C, D, 1, b, c, d
2, 3, i, j(a hole)
6, H, E, F, 7, f (a hole)

Hill - Chapter 4

09/23/99

page 59

POLYA - POLYB:
4, 5, 6, H, E, F, 7, e, 8, B, C, D, 1, a
2, 3, k
POLYB - POLYA:
1, b, c, d, 8, 5, g, h, 4, A, 3, i, j, 2
7, f, 6, G
1st edition Figure A6.8.
Figure 4.64. Forming the union and difference of two polygons.
Notice how the holes (E, F, G, H) and (k, i, j) in the polygons are properly handled, and that the algorithm
generates holes as needed (holes are polygons listed in counterclockwise fashion).
Task: Adapt the WeilerAtherton method so that it can form the union and difference of two polygons, and
exercise your routines on a variety of polygons. Generate A and B polygons, either in files or
algorithmically, to assist in the testing. Draw the polygons A and B in two different colors, and the result of
the operation in a third color.

4.11. For Further Reading

Many books provide a good introduction to vectors. A favorite is Hoffmanns ABOUT VECTORS. The
GRAPHICS GEMS series [gems] provides an excellent source of new approaches and results in vector
arithmetic and geometric algorithms by computer graphics practitioners. Three excellent example articles
are Alan Paeths A Half-Angle Identity for Digital Computation: The Joys of the Half Tangent
[paeth91], Ron Goldman Triangles[goldman90], and Lopez-Lopezs Triangles Revisited [lopez92].
Two books that delve more deeply into the nature of geometric algorithms are Moret and Shapiros
ALGORITHMS FROM P TO NP [moret91] and Preparata and Shamoss COMPUTATIONAL
GEOMETRY, AN INTRODUCTION [preparata85].

Hill - Chapter 4

09/23/99

page 60

Chapter 5. Transformations of Objects

Minus times minus is plus, the reason for this we need not discuss
W.H.Auden
If I eat one of these cakes, she thought,
it's sure to make some change in my size.
So she swallowed one
and was delighted to find that she began shrinking directly.
Lewis Carroll, Alice in Wonderland
Many of the brightly colored tile-covered walls and floors of the
Alhambra in Spain show us that the Moors were masters
in the art of filling a plane with similar interlocking figures,
bordering each other without gaps.
What a pity that their religion forbade them to make images!
M. C. Escher

Goals of the Chapter

Develop tools for transforming one picture into another.

Introduce the fundamental concepts of affine transformations, which perform combinations of rotations,
scalings, and translations.
Develop functions that apply affine transformations to objects in computer programs
Develop tools for transforming coordinate frames.
See how to set up a camera to render a 3D scene using OpenGL
Learn to design scenes in the Scene Design language SDL, and to write programs that read SDL files and
draw the scenes they describe.

Preview.
Section 5.1 motivates the use of 2D and 3D transformations in computer graphics, and sets up some basic
definitions. Section 5.2 defines 2D affine transformations and establishes terminology for them in terms of a
matrix. The notation of coordinate frames is used to keep clear what objects are being altered and how. The
section shows how elementary affine transformations can perform scaling, rotation, translation, and shearing.
Section 5.2.5 shows that you can combine as many affine transformations as you wish, and the result is another
affine transformation, also characterized by a matrix. Section 5.2.7 discusses key properties of all affine
transformations most notably that they preserve straight lines, planes, and parallelism and shows why they
are so prevalent in computer graphics.
Section 5.3 extends these ideas to 3D affine transformations, and shows that all of the basic properties hold here
as well. 3D transformations are more complex than 2D ones, however, and more difficult to visualize,
particularly when it comes to 3D rotations. So special attention is paid to describing and combining various
rotations.
Section 5.4 discusses the relationship between transforming points and transforming coordinate systems. Section
5.5 shows how transformations are managed within a program when OpenGL is available, and how
transformations can greatly simplify many operations commonly needed in a graphics program. Modeling
transformations and the use of the current transformation are motivated through a number of examples.
Section 5.6 discusses modeling 3D scenes and drawing them using OpenGL. A camera is defined that is
positioned and oriented so that it takes the desired snapshot of the scene. The section discusses how
transformations are used to size and position objects as desired in a scene. Some example 3D scenes are modeled
and rendered, and the code required to do it is examined. This section also introduces a Scene Description
language, SDL, and shows how to write an application that can draw any scene described in the language. This

Chap 5. Transformations

9/28/99

page 1

requires the development of a number of classes to support reading and parsing SDL files, and creating lists of
objects that can be rendered. These classes are available from the books web site.
The chapter ends with a number of Case Studies that elaborate on the main ideas and provide opportunities to
work with affine transformations in graphics programs. One case study asks you to develop routines that perform
transformations when OpenGL is not available. Also described there are ways to decompose an affine
transformation into its elementary operations, and the development of a fast routine to draw arcs of circles that
capitalizes on the equivalence between a rotation and three successive shears.

5.1. Introduction.
The main goal in this chapter is to develop techniques for working with a particularly powerful family of
transformations called affine transformations, both with pencil and paper and in a computer program, with and
without OpenGL. These transformations are a fundamental cornerstone of computer graphics, and are central to
OpenGL as well as most other graphics systems. They are also a source of difficulty for many programmers
because it is often difficult to get them right.
One particularly delicate area is the confusion of points and vectors. Points and vectors seem very similar, and are
often expressed in a program using the same data type, perhaps a list of three numbers like (3.0, 2.5, -1.145) to
express them in the current coordinate system. But this practice can lead to disaster in the form of serious bugs
that are very difficult to ferret out, principally because points and vectors do not transform the same way. We need
a way to keep them straight, which is offered by using coordinate frames and appropriate homogeneous
coordinates as introduced in Chapter 4.

5.2. Introduction to Transformations.

The universe is full of magical
things, patiently waiting for
our wits to grow sharper.
E. Phillpotts
We have already seen some examples of transformations, at least in 2D. In Chapter 3, for instance, the window to
viewport transformation was used to scale and translate objects situated in the world window to their final size and
position in the viewport.
We want to build on those ideas, and gain more flexible control over the size, orientation, and position of objects
of interest. In the following sections we develop the required tools, based on the powerful affine transformation,
which is a staple in computer graphics. We operate in both two and three dimensions.
Figure 5.1a shows two versions of a simple house, drawn before and after each of its points has been transformed.
In this case the house has been scaled down in size, rotated a small amount, and then moved up and to the right.
The overall transformation here is a combination of three more elementary ones: scaling, rotation, and translation.
Figure 5.1b shows a 3D house before and after it is similarly transformed: each 3D point in the house is subjected
by the transformation to a scaling, a rotation, and a translation.
a).

b).

aft er

aft er
before
before
x

Figure 5.1. Drawings of objects before and after they are transformed.

Chap 5. Transformations

9/28/99

page 2

Transformations are very useful in a number of situations:

a). We can compose a scene out of a number of objects, as in Figure 5.2. Each object such as the arch is most
easily designed (once for all) in its own master coordinate system. The scene is then fashioned by placing a
number of instances of the arch at different places and with different sizes, using the proper transformation for
each.

Figure 5.2. Composing a picture from many instances of a simple form.

Figure 5.3 shows a 3D example, where the scene is composed of many instances of cubes that have been scaled
and positioned in a city.

Figure 5.3. Composing a 3D scene from primitives.

b). Some objects, such as the snowflake shown in Figure 5.4, exhibit certain symmetries. We can design a single
motif and then fashion the whole shape by appropriate reflections, rotations, and translations of the motif.
use it 12 times

Figure 5.4. Using a motif to build up a figure.

c). A designer may want to view an object from different vantage points, and make a picture from each one. The
scene can be rotated and viewed with the same camera. But as suggested in Figure 5.5 its more natural to leave

Chap 5. Transformations

9/28/99

page 3

the scene alone and move the camera to different orientations and positions for each snapshot. Positioning and
reorienting a camera can be carried out through the use of 3D affine transformations.

scene

Figure 5.5. Viewing a scene from different points of view.

d). In a computer animation several objects must move relative to one another from frame to frame This can
be achieved by shifting and rotating their local coordinate systems as the animation proceeds. Figure 5.6
shows an example.
author supplied
Figure 5.6. Animating by transforming shapes.

Where are we headed? Using Transformations with OpenGL.

The first few sections of this chapter present the basic concepts of affine transformations, and show how they
produce certain geometric effects such as scaling, rotations, and translations, both in 2D and 3D space.
Ultimately, of course, the goal is to produce graphical drawings of objects that have been transformed to the
proper size, orientation, and position so they produce the desired scene. A number of graphics platforms,
including OpenGL, provide a graphics pipeline, or sequence of operations that are applied to all points that are
sent through it. A drawing is produced by processing each point.
Figure 5.7 shows a simplified view of the OpenGL graphics pipeline. An application sends it a sequence of
points P1, P2, P3, ... using the commands like the now familiar:

Figure 5.7. The OpenGL pipeline.

glBegin(GL_LINES);
glVertex3f(...); // send P1 through the pipeline
glVertex3f(...); // send P2 through the pipeline
glVertex3f(...); // send P3 through the pipeline
...
glEnd();
As shown in the figure these points first encounter a transformation called the current transformation (CT),
which alters their values into a different set of points, say Q1, Q2, Q3, .... Just as the original points Pi describe
some geometric object, the points Qi describe the transformed version of the same object. These points are then
sent through additional steps, and ultimately are used to draw the final image on the display.

Chap 5. Transformations

9/28/99

page 4

The current transformation therefore provides a crucial tool in the manipulation of graphical objects, and it is
essential for the application programmer to know how to adjust the CT so that the desired transformations are
produced. After developing the underlying theory of affine transformations, we turn in Section 5.5 to showing
how this is done.

Object transformations versus coordinate transformations.

There are two ways to view a transformation: as an object transformation or as a coordinate transformation.
An object transformation alters the coordinates of each point on the object according to some rule, leaving the
underlying coordinate system fixed. A coordinate transformation defines a new coordinate system in terms of the
old one, then represents all of the objects points in this new system. The two views are closely connected, and
each has its advantages, but they are implemented somewhat differently. We shall first develop the central ideas in
terms of object transformations, and then relate them to coordinate transformations.

5.2.1. Transforming Points and Objects.

We look here at the general idea of a transformation, and then specialize to affine transformations.
A transformation alters each point, P, in space (2D or 3D) into a new point, Q, using a specific formula or
algorithm. Figure 5.8 shows 2D and 3D examples.
y
a).

b).

y
T

Q
P

P
x

Figure 5.8. Mapping points into new points.

As Figure 5.8 illustrates, an arbitrary point P in the plane is mapped to Q. We say Q is the image of P under the
mapping T. Part a) shows a 2D point P being mapped to a new point Q; part b) shows a 3D point P being mapped
to a new Q. We transform an object by transforming each of its points, using, of course, the same function T() for
each point. We can map whole collections of points at once. The collection might be all the points on a line or
circle. The image of line L under T , for instance, consists of the images of all the individual points of L.1
Most mappings of interest are continuous, so the image of a straight line is still a connected curve of some shape,
although its not necessarily a straight line. Affine transformations, however, do preserve lines as we shall see: The
image under T of a straight line is also a straight line. Most of this chapter will focus on affine transformations, but
other kinds can be used to create special effects. Figure 5.9, for instance, shows a complex warping of a figure that
cannot be achieved with an affine transformation. This transformation might be used for visual effect, or to
emphasize important features of an object.
old Fig 11.2 Da Vinci warped peculiarly
Figure 5.9. A complex warping of a figure.
To keep things straight we use an explicit coordinate frame when performing transformations. Recall from
Chapter 4 that a coordinate frame consists of a particular point, , called the origin, and some mutually
perpendicular vectors (called i and j in the 2D case; i, j, and k in the 3D case) that serve as the axes of the
coordinate frame.

1More

formally, if S is a set of points, its image T(S) is the set of all points T(P) where P is some point in S.

Chap 5. Transformations

9/28/99

page 5

Take the 2D case first, as it is easier to visualize. In whichever coordinate frame we are using, point P and Q have

the representations P and Q given by:

Px
Qx
~
~
P = Py , Q = Qy
1
1

Recall that this means the point P is at location P = Px i + Py j + , and similarly for Q. Px and Py are
~
familiarly called the coordinates of P. The transformation operates on the representation P and produces the

representation Q according to some function, T(),

Q P
Q = T P
1 1
x

(5.1)

or more succinctly,

~
~
Q = T ( P) .

(5.2)

The function T() could be complicated, as in

Q cos(ln(PP)e)
Q = 1 +
1 1P
x

and such transformations might have interesting geometric effects, but we restrict ourselves to much simpler
families of functions, those that are linear in Px and Py This property characterizes the affine transformations.

5.2.2. The Affine Transformations.

What is algebra, exactly? Is it those three-cornered things?
J. M. Barrie
Affine transformations are the most common transformations used in computer graphics. Among other things they
make it easy to scale, rotate, and reposition figures. A succession of affine transformations can easily be combined
into a simple overall affine transformation, and affine transformations permit a compact matrix representation.
Affine transformations have a simple form: the coordinates of Q are linear combinations of those of P:

Q m P + m P + m
Q = m P + m P + m
1
1
x

11 x

12 y

21 x

22 y

(5.3)

for some six given constants m11, m12, etc. Qx consists of portions of both of Px and Py, and so does Qy. This cross
fertilization between the x- and y-components gives rise to rotations and shears.

Chap 5. Transformations

9/28/99

page 6

The affine transformation of Equation 5.3 has a useful matrix representation that helps to organize your thinking:2

Q m
Q = m
1 0
x

m12
m22

P
P
1 1

m13
m23

(5.4)

(Just multiply this out to see that its the same as Equation 5.3. In particular, note how the third row of the matrix
forces the third component of Q to be 1.) For an affine transformation the third row of the matrix is always (0, 0, 1).
Vectors can be transformed as well as points. Recall that if vector V has coordinates Vx and Vy then its coordinate
frame representation is a column vector with a third component of 0. When transformed by the same affine
transformation as above the result is

W m
W = m
0 0
x

V
V
1 0

m12
m22

m13
m23

(5.5)

which is clearly another vector: its third component is always 0.

3
2
0

Practice Exercise 5.2.1. Apply the transformation. An affine transformation is specified by the matrix:

0 5
1 2 .
0 1

8 3
Solution: 2 = 2
1 0

Find the image Q of point P = (1, 2).

0 5 1
1 2 2 .
0 1 1

5.2.3. Geometric Effects of Elementary 2D Affine Transformations.

What geometric effects are produced by affine transformations? They produce combinations of four elementary
transformations: (1) a translation, (2) a scaling, (3) a rotation, and (4) a shear.
Figure 5.10 shows an example of the effect of each kind of transformation applied individually.
1st Ed. Figure 11.5
Figure 5.10. Transformations of a map: a) translation, b) scaling, c)rotation, d) shear.

Translation.
You often want to translate a picture into a different position on a graphics display. The translation part of the
affine transformation arises from the third column of the matrix

Q 1
Q = 0
1 0
x
y

2See

0 m13
1 m23
0 1

P
P
1
x

(5.6)

Appendix 2 for a review of matrices.

Chap 5. Transformations

9/28/99

page 7

or simply

Q P + m
Q = P + m
1 1
x

so in ordinary coordinates Q = P + d, where offset vector d has components (m13, m23).

For example, if the offset vector is (2, 3), every point will be altered into a new point that is two units farther to the
right and three units above the original point. The point (1, -5), for instance, is transformed into (3, - 2), and the
point (0, 0) is transformed into (2, 3).

Scaling.
A scaling changes the size of a picture and involves two scale factors, Sx and Sy, for the x- and y-coordinates,
respectively:
(Qx, Qy) = (SxPx, SyPy)
Thus the matrix for a scaling by itself is simply

S
0
0

0
Sy
0

0
0
1

(5.7)

Scaling in this fashion is more accurately called scaling about the origin, because each point P is moved Sx times
farther from the origin in the x-direction, and Sy times farther from the origin in the y-direction. If a scale factor is
negative, then there is also a reflection about a coordinate axis. Figure 5.11 shows an example in which the scaling
(Sx, Sy) = (-1, 2) is applied to a collection of points. Each point is both reflected about the y-axis and scaled by 2 in
the y-direction.
y

Figure 5.11. A scaling and a reflection.

There are also pure reflections, for which each of the scale factors is +1 or -1. An example is
T(Px, Py) = (-Px, Py)

(5.8)

which produces a mirror image of a picture by flipping it horizontally about the y-axis, replacing each
occurrence of x with -x. (What is the matrix of this transformation?)
If the two scale factors are the same, Sx = Sy = S, the transformation is a uniform scaling, or a magnification
about the origin, with magnification factor |S|. If S is negative, there are reflections about both axes. A point is

Chap 5. Transformations

9/28/99

page 8

moved outward from the origin to a position |S| times farther away from the origin. If |S| < 1, the points will be
moved closer to the origin, producing a reduction (or demagnification). If, on the other hand, the scale factors
are not the same, the scaling is called a differential scaling.
Practice Exercise 5.2.2. Sketch the effect. A pure scaling affine transformation uses scale factors Sx = 3 and Sy =
-2. Find the image of each of the three objects in Figure 5.12 under this transformation, and sketch them. (Make
use of the facts - to be proved later - that an affine transformations maps straight lines to straight lines, and ellipses
to ellipses.)
a).
b).
c).

1
(1, 1)
1

Figure 5.12. Objects to be scaled.

Rotation.
A fundamental graphics operation is the rotation of a figure about a given point through some angle. Figure 5.13
o
shows a set of points rotated about the origin through an angle of = 60 .
y

x
60
T

Figure 5.13. Rotation of points through an angle of 60.

When T( ) is a rotation about the origin, the offset vector d is zero and Q = T(P) has the form

Qx = Px cos( ) Py sin( )
(5.9)

Qy = Px sin( ) + Py cos( )

As we derive next, this form causes positive values of to perform a counterclockwise (CCW) rotation. In terms
of its matrix form, a pure rotation about the origin is given by

cos( )
sin( )
0

sin( ) 0
cos( ) 0
0
1

(5.10)

Example 5.2.1. Find the transformed point, Q, caused by rotating P = (3, 5) about the origin through an angle of
0
0
60 . Solution: For an angle of 60 , cos() = .5 and sin() = .866, and Equation 5.9 yields Qx = (3)(0.5) 0

(5)(0.866) = -2.83 and Qy = (3)(0.866) + (5)(0.5) = 5.098. Check this on graph paper by swinging an arc of 60
from (3, 5) and reading off the position of the mapped point. Also check numerically that Q and P are at the same
distance from the origin. (What is this distance?)

Chap 5. Transformations

9/28/99

page 9

Derivation of the Rotation Mapping.

We wish to demonstrate that Equation 5.9 is correct. Figure 5.14 shows how to find the coordinates of a point Q
that results from rotating point P about the origin through an angle . If P is at a distance R from the origin, at
some angle , then P = (R cos( ), R sin( )). Now Q must be at the same distance as P , and at angle + . Using
trigonometry, the coordinates of Q are
y

P
R

Figure 5.14. Derivation of the rotation mapping.

Qx = R cos( + )
Qy = Rsin( + )
Substitute into this equation the two familiar trigonometric relations:
cos( + ) = cos() cos() - sin() sin()
sin(+ ) = sin() cos() + cos() sin()
and use Px = R cos() and Py = R sin() to obtain Equation 5.9.
Practice Exercise 5.2.3. Rotate a Point. Use Equation 5.9 to find the image of each of the following points after
rotation about the origin:
a). (2, 3) through an angle of - 45o
b). (1, 1) through an angle of - 180 o.
c). (60, 61) through an angle of 4 o.
In each case check the result on graph paper, and compare numerically the distances of the original point and its
image from the origin.
Solution: a). (3.5355, .7071), b). (-1, -1), c). (55.5987, 65.0368).

Shearing.
An example of shearing is illustrated in Figure 5.15 is a shear in the x-direction (or along x). In this case the ycoordinate of each point is unaffected, whereas each x-coordinate is translated by an amount that increases linearly
with y. A shear in the x-direction is given by
y
before

after

Figure 5.15. An example of shearing.

Qx = Px + h Py

Chap 5. Transformations

9/28/99

page 10

Qy = Py
where the coefficient h specifies what fraction of the y-coordinate of P is to be added to the x-coordinate. The
quantity h can be positive or negative. Shearing is sometimes used to make italic letters out of regular letters. The
matrix associated with this shear is:

1
0
0

h 0
1 0
0 1

(5.11)

One can also have a shear along y, for which Qx = Px and Qy = g Px + Py for some value g, so that the matrix is
given by

1
g
0

0 0
1 0
0 1

(5.12)

Example 5.2.2: Into which point does (3, 4) shear when h = .3 in Equation 5.11? Solution: Q = (3 + (.3)4 , 4) = (4.2, 4).
Example 5.2.3: Let g = 0.2 in Equation 5.12. To what point does (6, - 2) map? Solution: Q = (6, 0.2 6 - 2) = (6, - 0.8).
A more general shear along an arbitrary line is discussed in a Case Study at the end of the chapter. A notable
feature of a shear is that its matrix has a determinant of 1. As we see later this implies that the area of a figure is
unchanged when it is sheared.
Practice Exercise 5.2.4. Shearing Lines. Consider the shear for which g = .4 and h = 0. Experiment with various
sets of three collinear points to build some assurance that the sheared points are still collinear. Then, assuming that
lines do shear into lines, determine into what objects the following line segments shear:
a. the horizontal segment between ( - 3, 4) and (2, 4);
b. the horizontal segment between ( - 3, - 4) and (2, - 4);
c. the vertical segment between ( - 2, 5) and ( - 2, - 1);
d. the vertical segment between (2, 5) and (2, - 1);
e. the segment between ( - 1, - 2) and (3, 2);
Into what shapes do each of the objects in Figure 5.2.12 shear?

5.2.4. The Inverse of an Affine Transformation

Most affine transformations of interest are nonsingular, which means that the determinant of M in Equation 5.4,
3

which evaluates to

det M = m11 m22 m12 m21

(5.13)

is nonzero. Notice that the third column of M, which represents the amount of translation, does not affect the
determinant. This is a direct consequence of the two zeroes appearing in the third row of M. We shall make special
note on those rare occasions that we use singular transformations.

3 See Appendix 2 for a review of determinants.

Chap 5. Transformations

9/28/99

page 11

It is reassuring to be able to undo the effect of a transformation. This is particularly easy to do with nonsingular
affine transformations. If point P is mapped into point Q according to Q = MP, simply premultiply both sides by
the inverse of M, denoted M -1, and write
P = M -1 Q

(5.14)

The inverse of M is given by4

m22 -m12
M -1 = 1
det M -m 21 m11

(5.15)

We therefore obtain the following matrices for the elementary inverse transformations:
Scaling (use M as found in Equation 5.7):

1
S
= 0
0

0

1

1
Sy
0

Rotation (use M as found in Equation 5.10):

cos( )
= sin( )
0

sin( ) 0
cos( ) 0
0
1

1
= h
0

Shearing (using the version of M in Equation 5.11):

0 0
1 0
0 1

Translations: The inverse transformation simply subtracts the offset rather than adds it.

1
= 0
0

0 m13
1 m23
0
1

Practice Exercises.
5.2.5. What Is the Inverse of a Rotation? Show that the inverse of a rotation through is a rotation through -.
Is this reasonable geometrically? Why?
5.2.6. Inverting a Shear. Is the inverse of a shear also a shear? Show why or why not.
5.2.7. An Inverse Matrix. Compute the inverse of the matrix

3
M = 1
0

2 1
1 0 .
0 1

See Appendix 2 for a review of inverse matrices.

Chap 5. Transformations

9/28/99

page 12

5.2.5. Composing Affine Transformations.

Progress might have been all right once , but it has gone on too long.
Ogden Nash
It's rare that we want to perform just one elementary transformation; usually an application requires that we build
a compound transformation out of several elementary ones. For example, we may want to
translate by (3, - 4)
then rotate through 300
then scale by (2, - 1)
then translate by (0, 1.5)
and finally rotate through - 300.
How do these individual transformations combine into one overall transformation? The process of applying
several transformations in succession to form one overall transformation is called composing (or concatenating)
the transformations. As we shall see, when two affine transformations are composed, the resulting transformation
is (happily) also affine.
Consider what happens when two 2D transformations, T1( ) and T2( ), are composed. As suggested in Figure 5.16,
T1( ) maps P into Q, and T2( ) maps Q into point W. What is the transformation, T( ), that maps P directly into W?
That is, what is the nature of W = T2(Q) = T2(T1(P))?

T 1()

T 2()

P
W
T()

Figure 5.16. The composition of two transformations.

and M . Thus P is first transformed to the

Suppose the two transformations are represented by the matrices M
1
2

P which is then transformed to M ( M P ). By associativity this is just ( M M ) P , and so we have

point M
1
2
1
2
1
~ ~~
W = MP

(5.16)

where the overall transformation is represented by the single matrix

M = M 2 M 1 .

(5.17)

When homogeneous coordinates are used, composing affine transformations is accomplished by simple matrix
multiplication. Notice that the matrices appear in reverse order to that in which the transformations are applied: if

we first apply T1 with matrix M1 , and then apply T2 with matrix M 2 to the result, the overall transformation has

~ ~

matrix M2 M1 , with the second matrix appearing first in the product as you read from left to right. (Just the
opposite order will be seen when we transform coordinate systems.)
By applying the same reasoning, any number of affine transformations can be composed simply by multiplying
their associated matrices. In this way, transformations based on an arbitrary succession of rotations, scalings,
shears, and translations can be formed and captured in a single matrix.

Chap 5. Transformations

9/28/99

page 13

Example 5.2.4. Build one. Build a transformation that

a). rotates through 45 degrees;
b). then scales in x by 1.5 and in y by - 2;
c). then translates through (3, 5).
Find the image under this transformation of the point (1,2).
Solution: Construct the three matrices and multiply them in the proper order (first one last, etc.) to form:

1
0
0

1.06 3
0 3 1.5 0 0 .707 .707 0
1.06
1 5 0 2 0 .707 .707 0 = 1.414 1.414 5
0 1 0 0 1
0
0
1
0
0
1

Now to transform point (1, 2), enlarge it to the triple (1, 2, 1), multiply it by the composite matrix to obtain (1.94,
0.758, 1), and drop the one to form the image point (1.94, 0.758). It is instructive to use graph paper, and to
perform each of this transformations in turn to see how (1, 2) is mapped.

5.2.6. Examples of Composing 2D Transformations.

Art is the imposing of a pattern on experience,
and our aesthetic enjoyment is recognition of the pattern.
Alfred North Whitehead
We examine some important examples of composing 2D transformations, and see how they behave.
Example 5.2.5. Rotating About an Arbitrary Point.
So far all rotations have been about the origin. But suppose we wish instead to rotate points about some other
point in the plane. As suggested in Figure 5.17, the desired pivot point is V = (Vx, Vy), and we wish to rotate
points such as P through angle to position Q. To do this we must relate the rotation about V to an elementary
rotation about the origin.
Q
Q'

V
P'

Figure 5.17. Rotation about a point.

Figure 5.2.17 shows that if we first translate all points so that V coincides with the origin, then a rotation about the
origin (which maps P to Q) will be appropriate. Once done, the whole plane is shifted back to restore V to its
original location. The rotation therefore consists of the following three elementary transformations:
1.
2.
3.

Translate point P through vector v = (-Vx, -Vy):

Rotate about the origin through angle ;
Translate P back through v.

Creating a matrix for each elementary transformation, and multiplying them out produces:

1
0
0

0 Vx
1 Vy
0 1

cos( )
sin( )
0

Chap 5. Transformations

sin( ) 0 1 0 Vx
cos( ) sin( ) d x
cos( ) 0 0 1 Vy = sin( ) cos( ) d y
0
1 0 0 1
0
0
1

9/28/99

page 14

where the overall translation components are

d x = cos( )Vx + sin( )Vy + Vx

d y = sin( )Vx cos( )Vy + Vy
Because the same cos() and sin() terms appear in this result as in a rotation about the origin, we see that a
rotation about an arbitrary point is equivalent to a rotation about the origin followed by a complicated translation
through (dx, dy).
As a specific example, we find the transformation that rotates points through 30 about (-2, 3), and determine to
which point the point (1, 2) maps. A 30 rotation uses cos() = 0.866 and sin() = 0.5. The offset vector is then
(1.232, 1.402), and so the transformation applied to any point (Px, Py) is
Qx = 0.866 Px - 0.5 Py + 1.232
Qy = 0.5 Px + 0.866 Py + 1.402
Applying this to (1, 2) yields (1.098, 3.634). This is the correct result, as can be checked by sketching it on graph
paper. (Do it!)
Example 5.2.6. Scaling and Shearing about arbitrary pivot points.
In a similar manner we often want to scale all points about some pivot point other than the origin. Because the
elementary scaling operation of Equation 5.13 scales points about the origin, we do the same shift-transformunshift sequence as for rotations. This and generalizing the shearing operation are explored in the exercises.
Example 5.2.7. Reflections about a tilted line.
Consider the line through the origin that makes an angle of with the x-axis, as shown in Figure 5.18. Point A
reflects into point B, and each house shown reflects into the other. We want to develop the transformation that
reflects any point P about this axis, to produce point Q. Is this an affine transformation?

A
P
b

ctio
refle

n ax

Figure 5.18. Reflecting about a tilted axis.

To show that it is affine, we build it out of three parts:
A rotation through angle - (so the axis coincides with the x-axis);
A reflection about the x-axis;
A rotation back through that restores the axis.
Each of these is represented by a matrix, so the overall transformation is given by the product of the three
matrices, so it is affine. Check that each of the steps is properly represented in the following three matrices, and
that the product is also correct:

Chap 5. Transformations

9/28/99

page 15

c
s
0

s 0 1 0 0 c s 0
c2 s2
c 0 0 1 0 s c 0 = 2cs
0 1 0 0 1 0 0 1
0

2cs
s2 c2
0

1
0
0

where c stands for cos() and s for sin(). Using trigonometric identities, the final matrix can be written (check
this out!)

cos(2 )
sin(2 )
0

sin( 2 ) 0
cos(2 ) 0
0
1

{a reflection about the axis at angle }

(5.18)

This has the general look of a rotation matrix, except the angle has been doubled and minus signs have crept into
the second column. But in fact it is the matrix for a reflection about the axis at angle .
Practice Exercises.
Exercise 5.2.8. The classic: the Window to Viewport Transformation.
We developed this transformation in Chapter 3. Rewriting Equation 3.2 in the current notation we have:

A 0 C
~
M= 0 B D
0 0 1

where the ingredients A, B, C, and D depend on the window and viewport and are given in Equation 3.3. Show
that this transformation is composed of:
A translation through (-W.l, -W.b) to place the lower left corner of the window at the origin;
A scaling by (A, B) to size things.
A translation through (V.l, V.b) to move the corner of the viewport to the desired position.
5.2.9. Alternative Form for a Rotation About a Point. Show that the transformation of Figure 5.17 can be
written out as
Qx = cos()(Px - Vx) - sin()(Py - Vy) + Vx
Qy = sin()(Px - Vx) + cos()(Py - Vy) + Vy
This form clearly reveals that the point is first translated by (-Vx, -Vy), rotated, and then translated by (Vx, Vy).
5.2.10. Where does it end up? Where is the point (8, 9) after it is rotated through 500 about the point (3, 1)? Find
the M matrix.
5.2.11. Seeing it two ways. On graph paper place point P = (4, 7) and the result Q of rotating P about V = (5, 4)
through 450. Now rotate P about the origin through 450 to produce Q, which is clearly different from Q. The
difference between them is V - VM. Show the point V - VM in the graph, and check that Q - Q equals V - VM.
5.2.12. What if the axis doesnt go through the origin? Find the affine transformation that produces a reflection
about the line given parametrically by L(t) = A + bt. Show that it reduces to the result in Equation 5.20 when A +
bt does pass through the origin.
5.2.13. Reflection in x = y. Show that a reflection about the line x = y is equivalent to a reflection in x followed
0
by a 90 rotation.
5.2.14. Scaling About an Arbitrary Point. Fashion the affine transformation that scales points about a pivot
point, (Vx, Vy). Test the overall transformation on some sample points, to confirm that the scaling operation is
correct. Compare this with the transformation for rotation about a pivot point.
5.2.15. Shearing Along a Tilted Axis. Fashion the transformation that shears a point along the axis described by
vector u tilted at angle , as shown in Figure 5.19. Point P is shifted along u an amount that is fraction f of the
displacement d of P from the axis.

Chap 5. Transformations

9/28/99

page 16

Q
P
d

Figure 5.19. Shearing along a tilted axis.

5.2.16. Transforming Three Points. An affine transformation is completely determined by specifying what it
does to three points. To illustrate this, find the affine transformation that converts triangle C with vertices ( - 3, 3),
(0, 3), and (0, 5) into equilateral triangle D with vertices (0, 0), (2, 0), and (1, 3 ), as shown in Figure 5.20.
y
(0, 5)
C

c
(-3, 3)

(0, 3)
(1,

x
(2, 0)

Figure 5.20. Converting one triangle into another.

Do this by a sequence of three elementary transformations:
1. Translate C down by 3 and right by 3 to place vertex c at c.
2. Scale in x by 2/3 and in y by 3 /2 so C matches D in width and height.
3. Shear by 1/ 3 in the x-direction to align the top vertex of C with that of D.
Check that this transformation does in fact transform triangle C into triangle D. Also find the inverse of this
transformation and show that it converts triangle D back into triangle C.
5.2.17. Fixed Points of an Affine Transformation. The point F is a fixed point of the affine transformation T(p)
= Mp if T(F) = F. That is, if F satisfies F M = F.
a). Show that when the third column of M is (0,0,1) - such that there is no translation, the origin is always a fixed
point of T .
b). Show that if F is a fixed point of T then for any P we have T(P) = M(P - F)+ F.
c). Show that F must always satisfy: F = d(I - M)-1. Does every affine transformation have a fixed point?
d). What is a fixed point for a rotation about point V? Show that it satisfies the relationship in part b above.
e). What is the fixed point for a scaling with scale factors Sx and Sy, about point V?
f). Consider the 5-th iterate of T() applied to P, given by R = T(T(T(T(T(P))))). (Recall IFSs described at the
end of Chapter 1). Use the result in b) to show a simple form for output R in terms of fixed point F of T(): R = (P
- F)M5 + F.
5.2.18. Finding Matrices. Give the explicit form of the 3-by-3 matrix representing each of the following
transformations:
a. Scaling by a factor of 2 in the x-direction and then rotating about (2, 1).
b. Scaling about (2, 3) and following by translation through (1, 1).
c. Shearing in x by 30%, scaling by 2 in x, and then rotating about (1, 1) through 30.
5.2.19. Normalizing a Box. Find the affine transformation that maps the box with corners (0, 0), (2, 1), (0, 5), and
( - 2, 4) into the square with corners (0, 0), (1, 0), (1, 1), and (0, 1). Sketch the boxes.

Chap 5. Transformations

9/28/99

page 17

5.2.20. Some Transformations Commute. Show that uniform scaling commutes with rotation, in that the
resulting transformation does not depend on the order in which the individual transformations are applied. Show
that two translations commute, as do two scalings. Show that differential scaling does not commute with rotation.
5.2.21. Reflection plus a rotation. Show that a reflection in x followed by a reflection in y is the same as a rotation
by 1800.
5.2.22. Two Successive Rotations. Suppose that R() denotes the transformation that produces a rotation through
angle . Show that applying R(1) followed by R(2) is equivalent to applying the single rotation R(1 + 2). Thus
successive rotations are additive.
5.2.23. A Succession of Shears. Find the composition of a pure shear along the x-axis followed by a pure shear
along the y-axis. Is this still a shear? Sketch by hand an example of what happens to a square centered at the origin
when subjected to a simultaneous shear versus a succession of shears along the two axes.

5.2.8. Some Useful Properties of Affine Transformations.

We have seen how to represent 2D affine transformations with matrices, how to compose complex
transformations from a sequence of elementary ones, and the geometric effect of different 2D affine
transformations. Before moving on to 3D transformations it is useful to summarize some general properties of
affine transformations. These properties are easy to establish, and because no reference is made of the
dimensionality of the objects being transformed, they apply equally well to 3D affine transformations. The only
fact about 3D transformations we need at this point is that, like their 2D counterparts, they can be represented in
homogeneous coordinates by a matrix.

1). Affine transformations preserve affine combinations of points.

We know that an affine combination of two points P1 and P2 is the point
W = a1 P1 + a2 P2,

where a1 + a2 = 1

What happens when we apply an affine transformation T() to this point W? We claim T(W) is the same affine
combination of the transformed points, that is:
Claim: T(a1 P1 + a2 P2) = a1 T(P1) + a2 T(P2),

(5.19)

For instance, T(0.7 (2, 9) + 0.3 (1, 6)) = 0.7 T((2, 9)) + 0.3 T((1, 6)).

The truth of this is simply a matter of linearity. Using homogeneous coordinates, the point T(W) is MW , and
we can do the following steps using linearity of matrix multiplication :

~~ ~ ~
~
~~
~~
MW = M (a1 P1 + a2 P2 ) = a1 MP1 + a2 MP2
which in ordinary coordinates is just a1T(P1) + a2T(P2) as claimed. The property that affine combinations of
points are preserved under affine transformations seems fairly elementary and abstract, but it turns out to be
pivotal. It is sometimes taken as the definition of what an affine transformation is.

2). Affine transformations preserve lines and planes.

Affine transformations preserve collinearity and flatness, and so the image of a straight line is another straight
line. To see this, recall that the parametric representation L(t) of a line through A and B is itself an affine
combination of A and B:
L(t) = (1 - t) A + t B
This is an affine combination of points, so by the previous result the image of L(t) is the same affine combination
of the images of A and B:
Q(t) = (1 - t) T(A) + t T(B),

Chap 5. Transformations

(5.20)

9/28/99

page 18

This is another straight line passing through T(A) and T(B). In computer graphics this vastly simplifies drawing
transformed line segments: We need only compute the two transformed endpoints T(A) and T(B) and then draw a
straight line between them! This saves having to transform each of the points along the line, which is obviously
impossible.
The argument is the same to show that a plane is transformed into another plane. Recall from Equation 4.45 that
the parametric representation for a plane can be written as an affine combination of points:
P(s, t) = sA + tB + (1 - s - t)C
When each point is transformed this becomes:
T(P(s, t)) = sT(A) + t T(B) + (1 - s - t)T(C)
which is clearly also the parametric representation of some plane.
Preservation of collinearity and flatness guarantees that polygons will transform into polygons, and planar
polygons (those whose vertices all lie in a plane) will transform into planar polygons. In particular, triangles will
transform into triangles.

3). Parallelism of lines and planes is preserved.

If two lines or planes are parallel, their images under an affine transformation are also parallel. This is easy to
show. We first do it for lines. Take an arbitrary line A + bt having direction b. It transforms to the line given in

~ ~ ~

homogeneous coordinates by M ( A + bt ) = MA + ( Mb)t which has direction vector Mb . This new direction
does not depend on point A. Thus two different lines A 1+ bt and A2 + bt that have the same direction will

transform into two lines both having the direction Mb , so they are parallel. An important consequence of this
property is that parallelograms map into other parallelograms.
The same argument applies to planes: its direction vectors (see Equation 4.43) transform into new direction
vectors whose values do not depend on the location of the plane. A consequence of this is that parallelepipeds5
map into other parallelepipeds.
Example 5.2.8. How is a grid transformed?
Because affine transformations map parallelograms into parallelograms they are rather limited in how much they
can alter the shape of geometrical objects. To illustrate this apply any 2D affine transformation T to a unit square
grid, as in Figure 5.21. Because a grid consists of two sets of parallel lines, T maps the square grid to another grid
consisting of two sets of parallel lines. Think of the grid carrying along whatever objects are defined in the
grid, to get an idea of how the objects are warped by the transformation. This is all that an affine transformation
can do: warp figures in the same way that one grid is mapped into another. The new lines can be tilted at any
angle; they can be any (fixed) distance apart; and the two new axes need not be perpendicular. And of course the
whole grid can be positioned anywhere in the plane.

j
O

v
O' u

Figure 5.21. A transformed grid.

5 As we see later, a parallelipiped is the 3D analog of a parallelogram: it has six sides that occur in pairs of parallel faces.

Chap 5. Transformations

9/28/99

page 19

The same result applies in 3D: all a 3D affine transformation can do is map a cubical grid into a grid of
parallelipipeds.

4). The Columns of the Matrix reveal the Transformed Coordinate Frame.
It is useful to examine the columns of the matrix M of an affine transformation, for they prescribe how the
coordinate frame is transformed. Suppose the matrix M is given by

m
M = m
0

11
21

m12
m22
0

m13
m23 = m1 m 2 m3
1

(5.21)

so its columns are m1, m2, and m3. The first two columns are vectors (their third component is 0) and the last
column is a point (its third component is a 1). As always the coordinate frame of interest is defined by the origin
, and the basis vectors i and j, which have representations:

0 1
0

= 0 , i = 0 ,and j = 1
1 0
0
Notice that vector i transforms into the vector m1 (check this out):

m1 = Mi
and similarly j maps into m2 and maps into the point m3. This is illustrated in Figure 5.22a. The coordinate
frame (i, j, ) transforms into the coordinate frame (m1, m2, m3), and these new objects are precisely the
columns of the matrix.
a).
b).
Figure 5.22. The transformation forms a new coordinate frame.
The axes of the new coordinate frame are not necessarily perpendicular, nor must they be unit length. (They are
still perpendicular if the transformation involves only rotations and uniform scalings.) Any point P = Pxi + Pyj +
transforms into Q = Px m1 + Pym2 + m3. It is sometimes very revealing to look at the matrix of an affine
transformation in this way.
Example 5.2.9. Rotation about a point. The transformation explored in Example 5.2.5 is a rotation of 30o about
the point (-2, 3). This yielded the matrix:

.866
.5
0

.5 1.232
.866 1.402
0
1

As shown in Figure 5.22b the coordinate frame therefore maps into the new coordinate frame with origin at
(1.232, 1.402, 1) and coordinate axes given by the vectors (0.866, 0.5, 0) and (-0.5, 0.866, 0). Note that these
axes are still perpendicular, since only a rotation is involved.

5). Relative Ratios Are Preserved.

Affine transformations have yet another useful property. Consider a point P that lies at the fraction t of the way
between two given points, A and B, as shown in Figure 5.23. Apply affine transformation T( ) to A , B, and P. We
claim the transformed point, T(P), also lies the same fraction t of the way between the images T(A) and T(B).
This is not hard to show (see the exercises).

Chap 5. Transformations

9/28/99

page 20

B
1-t
T
t
A

P
t
T(A)

1-t
T(P)

T(B)
Figure 5.23. Relative ratios are preserved.
As a special case, midpoints of lines map into midpoints. This result pops out a nice geometric result: the
diagonals of any parallelogram bisect each other. (Proof: any parallelogram is an affine-transformed square
(why?), and the diagonals of a square bisect each other, so the diagonals of a parallelogram also bisect each
other.) The same applies in 3D space: the diagonals of any parallelepiped bisect each other.
Interesting Aside. In addition to preserving lines, parallelism, and relative ratios, affine transformations also
preserve ellipses and ellipsoids, as we see in Chapter 8!

6). Effect of Transformations on the Areas of Figures.

In CAD applications it is often important to compute the area or volume of an object. For instance, how is the
area of a polygon affected when all of its vertices are subjected to an affine transformation? It is clear
geometrically that neither translations nor rotations have any effect on the area of a figure, but scalings certainly
do, and shearing might.
The result is simple, and is developed in the exercises: When the 2D transformation with matrix M is applied to
an object, its area is multiplied by the magnitude of the determinant of M:

area after transformation

= det M
area before transformation

(5.22)

In 2D the determinant of M in Equation 5.6 is m11m22 - m12m21.6 Thus for a pure scaling as in Equation 5.10, the
new area is SxSy times the original area, whereas for a shear along one axis the new area is the same as the
original area! Equation 5.21 also confirms that a rotation does not alter the area of a figure, since cos2() +
sin2() = 1.
In 3D similar arguments apply, and we can conclude that the volume of a 3D object is scaled by |det M| when the
object is transformed by the 3D transformation based on matrix M.
Example 5.2.9: The Area of an Ellipse. What is the area of the ellipse that fits inside a rectangle with width W
and height H? Solution: This ellipse can be formed by scaling the unit circle x2 + y2 = 1 by the scale factors Sx =
W and Sy = H, a transformation for which the matrix M has determinant WH. The unit circle is known to have
area , and so the ellipse has area WH.

7). Every Affine Transformation is Composed of Elementary Operations.

We can construct complex affine transformations by composing a number of elementary ones. It is interesting to
turn the question around and ask, what elementary operations reside in a given affine transformation?

6 The determinant of the homogeneous coordinate version is the same (why?)

Chap 5. Transformations

9/28/99

page 21

Basically a matrix M may be factored into a product of elementary matrices in various ways. One particular way
~
of factoring the matrix M associated with a 2D affine transformation, elaborated upon in Case Study 5.3, yields
the result:

~
M = (shear)(scaling)(rotation)(translation)
~

That is, any 3 by 3 matrix M that represents a 2D affine transformation can be written as the product of (reading
right to left) a translation matrix, a rotation matrix, a scaling matrix, and a shear matrix. The specific ingredients
of each matrix are given in the Case Study.

In 3D things are somewhat more complicated. The 4 by 4 matrix M that represents a 3D affine transformation
can be written as:

~
M = (scaling)(rotation)(shear1)(shear2)(translation),
the product of (reading right to left) a translation matrix, a shear matrix, another shear matrix, a rotation matrix,
and a scaling matrix. This result is developed in Case study 5.???.
Practice Exercises.
5.2.294 Generalizing the argument. Show that if W is an affine combination of the N points Pi, i = 1,..,N, and
T() is an affine transformation, then T(W) is the same affine combination of the N points T(Pi), i = 1,..,N.
5.2.25. Show that relative ratios are preserved. Consider P given by A + bt where b = B - A. Find the distances
|P - A| and |P - B| from P to A and B respectively, showing that they lie in the ratio t to 1 - t. Is this true if t lies
outside of the range 0 to 1? Do the same for the distances |T(P) - T(A)| and |T(P) - T(B)|.
5.2.26. Effect on Area. Show that a 2D affine transformation causes the area of a figure to be multiplied by the
factor given in Equation 5.27. Hint: View a geometric figure as made up of many very small squares, each of
which is mapped into a parallelogram, and then find the area of this parallelogram.

5.3. 3D Affine Transformations.

The same ideas apply to 3D affine transformations as apply to 2D affine transformations, but of course the
expressions are more complicated, and it is considerably harder to visualize the effect of a 3D transformation.
Again we use coordinate frames, and suppose that we have an origin and three mutually perpendicular axes in
the directions i, j, and k (see Figure 5.8). Point P in this frame is given by P = + Pxi + Pyj + Pzk, and so has as
the representation

P
~ P
P=
P
1
x

y
z

Suppose T() is an affine transformation that transforms point P to point Q. Then just as in the 2D case T() is
~
represented by a matrix M which is now 4 by 4:

m
~ m
M=
m
0

11
21
31

m12
m22
m32

m13
m23
m33

Chap 5. Transformations

m14
m24
m34

(5.23)

9/28/99

page 22

and we can say that the representation of point Q is found by multiplying P by matrix M :

Q P
Q = M~ P
Q P
1 1
x

(5.24)

Notice that once again for an affine transformation the final row of the matrix is a string of zeroes followed a lone
one. (This will cease to be the case when we examine projective matrices in Chapter 7.)

5.3.1. The Elementary 3D Transformations.

We consider the nature of elementary 3D transformations individually, and then compose them into general 3D
affine transformations.

Translation.

For a pure translation, the matrix M has the simple form.

1
0
0
0

0 0 m14
1 0 m24
0 1 m34
0 0 1

Check that Q = MP is simply a shift in Q by the vector m = (m14, m24, m34).

Scaling.
Scaling in three dimensions is a direct extension of the 2D case, having a matrix given by:

S
0
0
0

0
Sy
0
0

0
0
Sz
0

0
0
0
1

(5.25)

where the three constants Sx, Sy, and Sz cause scaling of the corresponding coordinates. Scaling is about the origin,
just as in the 2D case. Figure 5.24 shows the effect of scaling in the z-direction by 0.5 and in the x-direction by a
factor of two.

Chap 5. Transformations

9/28/99

page 23

Figure 5.24. Scaling the basic barn.

Notice that this figure shows various lines before and after being transformed. It capitalizes on the important fact
that straight lines transform to straight lines.

Shearing.
Three-dimensional shears appear in greater variety than do their two-dimensional counterparts. The matrix for the
simplest elementary shear is the identity matrix with one zero term replaced by some value, as in

1
f
0
0

0
1
0
0

0
0
1
0

0
0
0
1

(5.26)

which produces Q = (Px, f Px + Py, Pz); that is, Py is offset by some amount proportional to Px, and the other
components are unchanged. This causes an effect similar to that in 2D shown in Figure 5.15. Goldman
[goldman???] has developed a much more general form for a 3D shear, which is described in Case Study 5.???.

Rotations.
Rotations in three dimensions are common in graphics, for we often want to rotate an object or a camera in order
to obtain different views. There is a much greater variety of rotations in three than in two dimensions, since we
must specify an axis about which the rotation occurs, rather than just a single point. One helpful approach is to
decompose a rotation into a combination of simpler ones.
Elementary rotations about a coordinate axis.
The simplest rotation is a rotation about one of the coordinate axes. We call a rotation about the x-axis an x-roll,
a rotation about the y-axis a y-roll, and one about the z-axis a z-roll. We present individually the matrices that
produce an x-roll, a y-roll, and a z-roll. In each case the rotation is through an angle, , about the given axis. We
define positive angles using a looking inward convention:
Positive values of cause a counterclockwise (CCW) rotation about an axis as one looks inward from a point on
the positive axis toward the origin.

Chap 5. Transformations

9/28/99

page 24

The three basic positive rotations are illustrated in Figure 5.25.7

Q ''
Q

P''

Figure 5.25. Positive rotations about the three axes.

This formulation is also consistent with our notion of 2D rotations: a positive rotation in two dimensions is
equivalent to a z-roll as we look at the xy-plane from a point on the positive z-axis.
Notice what happens with this convention for the particular case of a 90 rotation:
For a z-roll, the x-axis rotates to the y-axis.
For an x-roll, the y-axis rotates to the z-axis.
For a y-roll, the z-axis rotates to the x-axis.
The following three matrices represent transformations that rotate points through angle about a coordinate axis.
We use the suggestive notation Rx( ), Ry( ), and Rz( ) to denote x-, y-, and z-rolls, respectively. The parameter is
the angle through which points are rotated, given in radians, and c stands for cos() and s for sin().
1. An x-roll:

1
0
R ( ) =
0
0
x

(5.27)

(5.28)

0 0
c s 0
s c 0
0 0 1

2. A y-roll:

c
0
R ( ) =
s
0
y

0
1
0
0

s 0
0 0
c 0
0 1

3. A z-roll:

7In a left-handed system the sense of a rotation through a positive would be CCW looking outward along the positive axis from the origin.
This formulation is used by some authors.

Chap 5. Transformations

9/28/99

page 25

c
s
R ( ) =
0
0
z

s 0 0
c 0 0
0 1 0
0 0 1

(5.29)

Note that 12 of the terms in each matrix are the zeros and ones of the identity matrix. They occur in the row and
column that correspond to the axis about which the rotation is being made (e.g., the first row and column for a xroll). They guarantee that the corresponding coordinate of the point being transformed will not be altered. The c
and s terms always appear in a rectangular pattern in the other rows and columns.
Aside: Why is the y-roll different? The -s term appears in the lower row for the x- and z-rolls, but in the upper
row for the y- roll. Is a y-roll inherently different in some way? This question is explored in the exercises.
Example 5.3.1. Rotating the barn. Figure 5.26 shows a barn in its original orientation (part a), and after
a -70 x-roll (part b), a 30 y-roll (part c), and a -90 z-roll (part d).
a). the barn
b). -700 x-roll

c). 300 y-roll

d). -900 z-roll

Figure 5.26. Rotating the basic barn.

Example 5.3.2. Rotate the point P = (3, 1, 4) through 30 about the y-axis. Solution: Using Equation 9.9.10 with
c = .866 and s = .5, P is transformed into

c
0
Q=
s
0

0 s 0 3
4.6
1 0 0 1
1
=
0 c 0 4
1.964
0 0 1 1
1

As expected, the y-coordinate of the point is not altered.

Chap 5. Transformations

9/28/99

page 26

Practice Exercises.
5.3.1. Visualizing the 90 Rotations. Draw a right-handed 3D system and convince yourself that a 90 rotation
(CCW looking toward the origin) about each axis rotates the other axes into one another, as specified in the
preceding list. What is the effect of rotating a point on the x-axis about the x-axis?
5.3.2. Rotating the Basic Barn. Sketch the basic barn after each vertex has experienced a 45 x-roll. Repeat for
y- and z-rolls.
5.3.3. Do a Rotation. Find the image Q of the point P = (1, 2, -1) after a 45 y-roll. Sketch P and Q in a 3D
coordinate system and show that your result is reasonable.
5.3.4. Testing 90 Rotations of the Axes. This exercise provides a useful trick for remembering the form of the
rotation matrices. Apply each of the three rotation matrices to each of the standard unit position vectors, i, j, and
k, using a 90 rotation. In each case discuss the effect of the transformation on the unit vector.
5.3.5. Is a y-roll indeed different? The minus sign in Equation 5.28 seems to be in the wrong place: on the
lower s rather than the upper one. Here you show that Equations 5.27-29 are in fact consistent. Its just a matter
of how things are ordered. Think of the three axes x, y, and z as occurring cyclically: x -> y -> z -> x -> y , etc.
If we are discussing a rotation about some current axis (x-, y-, or z-) then we can identify the previous axis
and the next axis. For instance, if x- is the current axis, then the previous one is z- and the next is y-. Show that
with this naming all three types of rotations use the same equations: Qcurr = Pcurr, Qnext = c Pnext - s Pprev, and Qprev
= s Pnext + c Pprev. Write these equations out for each of the three possible current axes.

5.3.2. Composing 3D Affine Transformations.

Not surprisingly, 3D affine transformations can be composed, and the result is another 3D affine transformation.
The thinking is exactly parallel to that which led to Equation 5.17 in the 2D case. The matrix that represents the
overall transformation is the product of the individual matrices M1 and M2 that perform the two transformations,
with M2 premultiplying M1:

M = M 2 M 1 .

(5.30)

Any number of affine transformations can be composed in this way, and a single matrix results that represents
the overall transformation.
Figure 5.27 shows an example, where a barn is first transformed using some M1, then that transformed barn is
again transformed using M2. The result is the same as the barn transformed once using M2M1.

Figure 5.27. Composing 3D affine transformations. (file: fig5.27.bmp)

5.3.3. Combining Rotations.

Results! Why, man, I have gotten a lot of results.

Chap 5. Transformations

9/28/99

page 27

I know several thousand things that wont work.

Thomas A. Edison
One of the most important distinctions between 2D and 3D transformations is the manner in which rotations
combine. In 2D two rotations, say R(1) and R(2), combine to produce R(1+2), and the order in which they are
combined makes no difference. In 3D the situation is much more complicated, because rotations can be about
different axes. The order in which two rotations about different axes are performed does matter: 3D rotation
matrices do not commute. We explore some properties of 3D rotations here, investigating different ways that a
rotation can be represented, and see how to create rotations that do a certain job.
Its very common to build a rotation in 3D by composing three elementary rotations: an x-roll followed by a yroll, and then a z-roll. Using the notation of Equations 5.27-29 for each individual roll, the overall rotation is
given by

M = Rx ( x )Ry ( y )Rz ( z )

(5.31)

In this context the angles 1, 2, and 3 are often called Euler8 angles. One form of Eulers Theorem asserts
that any 3D rotation can be obtained by three rolls about the x-, y-, and z-axes, so any rotation can be written as
in Equation 5.32 for the appropriate choice of Euler angles. This implies that it takes three values to completely
specify a rotation.
Example 5.3.3. What is the matrix associated with an x-roll of 450 followed by a y-roll of 300 followed by a zroll of 600 ? Direct multiplication of the three component matrices (in the proper reverse order) yields:

.5
.866
0
0

.866
.5
0
0

0
0
1
0

0 .866
0
0
0 .5
1
0

0 .5 .0
1
0
0
0 .866 0
0
0
1

1
0
0
0

.707 .707
.707 .707
0
0

.433
0 .75
=
0 .5

1 0
0

.436

.789

.66
.612
0

.047
.612
0

0

0

1
0

Some people use a different ordering of rolls to create a complicated rotation. For instance, they might express a
rotation as Ry ( 1 )Rz ( 2 )Rx ( 3 ) : first a y-roll then a z-roll then an x-roll. Because rotations in 3D do not
commute this requires the use of different Euler angles 1, 2, and 3 to create the same rotation as in Equation
5.32. There are 12 possible orderings of the three individual rolls, and each uses different values for 1, 2, and 3.
Rotations About an Arbitrary Axis.
When using Euler angles we perform a sequence of x-, y-, and z-rolls, that is, rotations about a coordinate axis.
But it can be much easier to work with rotations if we have a way to rotate about an axis that points in an
arbitrary direction. Visualize the earth, or a toy top, spinning about a tilted axis. In fact, Eulers theorem states
that every rotation can be represented as one of this type:
Eulers Theorem: Any rotation (or sequence of rotations) about a point is equivalent to a single rotation about
some axis through that point.9
What is the matrix for such a rotation, and can we work with it conveniently?
Figure 5.28 shows an axis represented by vector u, and an arbitrary point P that is to be rotated through angle
about u to produce point Q.

8 Leonhard Euler, 1707-1783, a Swiss mathematician of extraordinary ability who made important contributions to all branches of
mathematics.

9 This is sometimes stated: Given two rectangular coordinate systems with the same origin and arbitrary directions of axes, one can always
specify a line through the origin such that one coordinate system goes into the other by a rotation about this line. [gellert75]

Chap 5. Transformations

9/28/99

page 28

z
u

Figure 5.28. Rotation about an axis through the origin.

Because u can have any direction, it would seem at first glance to be very difficult to find a single matrix that
represents such a rotation. But in fact it can be found in two rather different ways, a classic way and a
constructive way.
1). The classic way. Decompose the required rotation into a sequence of known steps:
1. Perform two rotations so that u becomes aligned with the z-axis.
2. Do a z-roll through angle .
3. Undo the two alignment rotations to restore u to its original direction.
This is reminiscent of rotating about a point in two dimensions: The first step prepares the situation for a simpler
known operation; the simple operation is done; and finally the preparation step is undone. The result (discussed
in the exercises) is that the transformation requires the multiplication of five matrices:
Ru() = Rz( -) Ry( -) Rz() Ry() Rz()

(5.32)

each being a rotation about one of the coordinate axes. This is tedious to do by hand but is straightforward to
carry out in a program. However, expanding out the product gives little insight into how the ingredients go
together.
2). The constructive way. Using some vector tools we can obtain a more revealing expression for the matrix
Ru(). This approach has become popular recently, and versions of it are described by several authors in GEMS I
[glass90]. We adapt the derivation of Maillot [mail90].
Figure 5.29 shows the axis of rotation u, and we wish to express the operation of rotating point P through angle
into point Q. The method, spelled out in Case Study 5.5, effectively establishes a 2D coordinate system in the
plane of rotation as shown. This defines two orthogonal vectors a and b lying in the plane, and as shown in
Figure 5.29b point Q is expressed as a linear combination of them. The expression for Q involves dot products
and cross products of various ingredients in the problem. But because each of the terms is linear in the
coordinates of P, it can be rewritten as P times a matrix.

Chap 5. Transformations

9/28/99

page 29

a).

b).

z
u
a'
h

P
p

Figure 5.29. P rotates to Q in the plane of rotation.

The final result is the matrix:

c + (1 c)u
(1 c)u u + su
R ( ) =
(1 c)u u su
0
2

x z

(1 c)u yu x suz
c + (1 c)uy 2
(1 c)uy uz + su x

(1 c)uz ux + su y
(1 c)uz uy sux
c + (1 c )uz 2

0
0
0
1

(5.33)

where c= cos(), and s = sin(), and (ux, uy, uz) are the components of the unit vector u. This looks more
complicated than it is. In fact, as we see later, there is so much structure in the terms that, given an arbitrary
rotation matrix, we can find the specific axis and angle that produces the rotation (which proves Eulers
theorem).
As we see later, OpenGL provides a function to create a rotation about an arbitrary axis:
glRotated(angle, ux, uy, uz);
Example 5.3.4. Rotating about an axis. Find the matrix that produces a rotation through 450 about the
axis u = (111
, , ) / 3 = (0.577,0.577,0.577) . Solution: For a 450 rotation, c = s = 0.707, and filling in the
terms in Equation 5.33 we obtain:

.8047
.5058
R ( 45 ) =
.31
0
0

.31 .5058
.8047 .31
.5058 .8047
0
0

0
0
0
1

This has a determinant of 1 as expected. Figure 5.30 shows the basic barn, shifted away from the origin, before it
is rotated (dark), after a rotation through 22.5o (medium), and after a rotation of 45o (light).

Chap 5. Transformations

9/28/99

page 30

Figure 5.30. The basic barn rotated about axis u.

Finding the Axis and Angle of Rotation.
Eulers theorem guarantees that any rotation is equivalent to a rotation about some axis. It is useful, when
presented with some rotation matrix, to determine the specific axis and angle. That is, given values mij for the
matrix:

m
m
R ( ) =
m
0

11
21

m12
m22
m32
0

m13
m23
m33
0

0

0

1
0

extract the angle and the unit vector u.

This is surprisingly easy to do by examining Equation 5.33[watt92]. First note that the trace of Ru(), that is the
2
2
2
sum of the three diagonal elements, is 3c + (1 c)(ux + uy + uz ) = 1 + 2cos( ) . So we can solve for
cos() directly:

cos( ) = 12 (m11 + m 22 + m33 1) .

Take the arc cosine of this value to obtain , and use it to find s = sin() as well. Now see that pairs of elements
of the matrix combine to reveal the individual components of u:

m32 m23
2 sin( )
m m31
uy = 13
2 sin( )
m m12
uz = 21
2 sin( )
ux =

Chap 5. Transformations

(5.34)

9/28/99

page 31

Example 5.3.5. Find the axis and angle. Pretend you dont know the underlying axis and angle for the rotation
matrix in Example 5.3.3, and solve for it. The trace is 2.414, so cos() = 0.707, must be 450, and sin() =
0.707. Now calculate each of the terms in Equation 5.35: they all yield the value 0.577, so u = (1,1,1)/ 3 ,
just as we expected.
Practice Exercises.
5.3.6. Which ones commute? Consider two affine transformations T1 and T2. Is T1T2 the same as T2T1 when:
a). They are both pure translations? b). They are both scalings? c). They are both shears?
d). One is a rotation and one a translation? e). One is a rotation and one is a scaling?
f). One is a scaling and one is a shear?
5.3.7. Special cases of rotation about a general axis u. It always helps to see that a complicated result collapses
to a familiar one in special cases. Check that this happens in Equation 5.34 when u is itself
a). the x-axis, i; b). the y-axis, j; c). the z-axis, k.
5.3.8. Classic Approach to Rotation about an Axis. Here we suggest how to find the rotations that cause u to
become aligned with the z-axis. (See Appendix 2 for a review of spherical coordinates.) Suppose the direction of
u is given by the spherical coordinate angles and as indicated in Figure 5.27. Align u with the z-axis by a zroll through -: this swings u into the xz-plane to form the new axis, u' (sketch this). Use Equation 5.29 to obtain
Rz(-). Second, a y-roll through - completes the alignment process. With u aligned along the z-axis, do the
desired z-roll through angle , using Equation 5.29. Finally, the alignment rotations must be undone to restore
the axis to its original direction. Use the inverse matrices to Ry( -) and Rz( -), which are Ry() and Rz( ),
respectively. First undo the y-roll and then the z-roll. Finally, multiply these five elementary rotations to obtain
Equation 5.34. Work out the details, and apply them to find the matrix M that performs a rotation through angle
35 about the axis situated at =30 and = 45. Show that the final result is:

.877
~ .445
M=
.124
0

.366 .281
.842 .306
.396
.910
0
0

0
0
0
1

5.3.9. Orthogonal Matrices. A matrix is orthogonal if its columns are mutually orthogonal unit-length vectors.
Show that each of the three rotation matrices given in Equations 5.27-29 is orthogonal. What is the determinant
of an orthogonal matrix? An orthogonal matrix has a splendid property: Its inverse is identical to its transpose
(also see Appendix 2). Show why the orthogonality of the columns guarantees this. Find the inverse of each of
the three rotation matrices above, and show that the inverse of a rotation is simply a rotation in the opposite
direction.
5.3.10. The Matrix Is Orthogonal. Show that the complicated rotation matrix in Equation 5.34 is orthogonal .
5.3.11. Structure of a rotation matrix. Show that for a 3x3 rotation M the three rows are pair-wise orthogonal,
and the third is the cross product of first two.
5.3.12. What if the axis of rotation does not pass through the origin? If the axis does not pass through the
origin but instead is given by S + ut for some point S, then we must first translate to the origin through -S, apply
the appropriate rotation, and then translate back through S. Derive the overall matrix that results.

5.3.4. Summary of Properties of 3D Affine Transformations.

The properties noted for affine transformations in Section 5.2.8 apply, of course, to 3D affine transformations.
Stated in terms of any 3D affine transformation T(.) having matrix M, they are:

Affine transformations preserve affine combinations of points. If a + b = 1 then aP + bQ is a meaningful

3D point, and T(aP + bQ) = aT(P) + bT(Q).
Affine transformations preserve lines and planes. Straightness is preserved: The image T(L) of a line L in
3D space is another straight line; the image T(W) of a plane W in 3D space is another plane.
Parallelism of lines and planes is preserved. If W and Z are parallel lines (or planes), then T(W) and T(Z)
are also parallel.

Chap 5. Transformations

9/28/99

page 32

The Columns of the Matrix reveal the Transformed Coordinate Frame. If the columns of M are the
vectors m1, m2, m3, and the point m4, the transformation maps the frame (i, j, k, ) to the frame (m1, m2,
m3, m4).
Relative Ratios Are Preserved. If P is fraction f of the way from point A to point B, then T(P) is also
fraction f of the way from point T(A) to T(B).
Effect of Transformations on the Areas of Figures. If 3D object D has volume V, then its image T(D) has
volume |det M | V, where |det M| is the absolute value of the determinant of M.
Every Affine Transformation is Composed of Elementary Operations. A 3D affine transformation may
be decomposed into a composition of elementary transformations. This can be done in several ways.

5.4. Changing Coordinate Systems.

Theres another way to think about affine transformations. In many respect it is a more natural approach when
modeling a scene. Instead of viewing an affine transformation as producing a different point in a fixed coordinate
system, you think of it as producing a new coordinate system in which to represent points.
Aside: a word on notation: To make things fit better on the printed page, we shall sometimes use the notation

P
in place of P . (Also see Appendix 2.) The superscript T denotes the transpose, so we are
1
x

(Px, Py, 1)

simply writing the column vector as a transposed row vector.

Suppose we have a 2D coordinate frame #1 as shown in Figure 5.31, with origin and axes i and j. Further
suppose we have an affine transformation T(.) represented by matrix M. So T(.) transforms coordinate frame #1
into coordinate frame #2, with new origin = T(), and new axes i = T(i) and j = T(j).

Figure 5.31. Transforming a coordinate frame.

Now let P be a point with representation (c, d, 1)T in the new system #2. What are the values of a and b in its
representation (a, b, 1)T in the original system #1? The answer: just multiply (c, d, 1)T by M:

a c
b = M d
1 1

(5.35)

Summarizing: Suppose coordinate system #2 is formed from coordinate system #1 by the affine transformation
M. Further suppose that point P = (Px, Py, Pz,1) are the coordinates of a point P expressed in system #2. Then the
coordinates of P expressed in system #1 are MP.
This may seem obvious to some readers, but in case it doesnt, a derivation is developed in the exercises. This
result also holds for 3D systems, of course, and we use it extensively when calculating how 3D points are
transformed as they are passed down the graphics pipeline.
Example 5.4.1. Rotating a coordinate system. Consider again the transformation of Example 5.2.5 that rotates
points through 30o about the point (-2, 3). (See Figure 5.22.) This transformation maps the origin and axes i
and j into the system #2 as shown in that figure. Now consider the point P with coordinates (Px, Py, 1)T in the new
coordinate system. What are the coordinates of this point expressed in the original system #1? The answer is
simply MP. For instance, (1, 2, 1)T in the new system lies at M(1, 2, 1)T = (1.098, 3.634, 1)T in the original
system. (Sketch this in the figure.) Notice that the point (-2, 3, 1)T, the center of rotation of the transformation, is

Chap 5. Transformations

9/28/99

page 33

a fixed point of the transformation: M(2, 3, 1)T = (2, 3, 1)T. Thus if we take P = (-2, 3, 1)T in the new system, it
maps to (-2, 3, 1)T in the original system (check this visually).
Successive Changes in a Coordinate frame.
Now consider forming a transformation by making two successive changes of the coordinate system. What is the
overall effect? As suggested in Figure 5.32, system #1 is converted to system #2 by transformation T1(.), and
system #2 is then transformed to system #3 by transformation T2(.). Note that system #3 is transformed relative
to #2.
y

b
f

e
d

system #3

T2
c

system #2
system #1
x
a

Figure 5.32. Transforming a coordinate system twice.

Again the question is: if point P has representation (e, f,1)T with respect to system #3, what are its coordinates (a,
b,1)T with respect to the original system #1?
To answer this just work backwards and collect the effects of each transformation. In terms of system #2 the
point P has coordinates (c, d,1)T = M2(e, f,1)T. And in terms of system #1 the point (c, d,1)T has coordinates (a,
b,1)T = M1( c, d,1)T. Putting these together:

a c
e
b = M d = M M f
1 1
1
1

(5.36)

The essential point is that when determining the desired coordinates (a, b,1)T from (e, f,1)T we first apply M2 and
then M1, just the opposite order as when applying transformations to points.
We summarize this fact for the case of three successive transformations. The result generalizes immediately to
any number of transformations.
Transforming points. To apply a sequence of transformations T1(), T2(), T3() (in that order) to a point P, form the
matrix:

M = M 3 M 2 M1
Then P is transformed to MP. To compose each successive transformation Mi you must premultiply by Mi.
Transforming the coordinate system. To apply a sequence of transformations T1(), T2(), T3() (in that order) to
the coordinate system, form the matrix:

M = M1 M2 M3

Chap 5. Transformations

9/28/99

page 34

Then a point P expressed in the transformed system has coordinates MP in the original system. To compose
each additional transformation Mi you must postmultiply by Mi.
How OpenGL operates.
We shall see in the next section that OpenGL provides tools for successively applying transformations in order to
build up an overall current transformation. In fact OpenGL is organized to postmultiply each new
transformation matrix to combine it with the current transformation. Thus it will often seem more natural to the
modeler to think in terms of successively transforming the coordinate system involved, as the order in which
these transformations is carried out is the same as the order in which OpenGL computes them.
Practice Exercises.
5.4.1. How transforming a coordinate system relates to transforming a point.
We wish to show the result in Equation 5.35. To do this, show each of the following steps.
a). Show why the point P with representation (c, d, 1)T used in system #2 lies at ci + dj + .
b). We want to find where this point lies in system #1. Show that the representation (in system #1) of i is M(1, 0,
0)T, that of j is M(0, 1, 0)T, and that of is M(0, 0, 1).
c). Show that therefore the representation of the point ci + dj + is cM(1, 0, 0)T + dM(0, 1, 0)T+ M(0, 0, 1)T.
d). Show that this is the same as M(c, 0, 0)T+ M(0, d, 0)T + M(0, 0, 1) and that this is M(c, d, 1)T, as claimed.
5.4.2. Using elementary examples. Figure 5.33 shows the effect of four elementary transformations of a
coordinate system. In each case the original system with axes x and y is transformed into the new system with
axes x and y.
a). translate(m, n)
b). rotate(a)
y'
y

(e,f)

(e,f)
x'

n
e

e+m

m
c). scale(3, 2)

d). scale(1, -1)

(e, f)

-f
6
4
2

3
2

x
x'

(2, 2)

1 1

Figure 5.33. Elementary changes between coordinate systems.

a). Part a) shows the effect of a translation through (m, n). Show that point (e, f) in the new system lies at (e + m,
f + n) in the original system.
b). Part b) shows the effect of a rotation about the origin through A degrees. Show that the point (e, f) in the new
system lies at (e cos(a) - f sin(a), e sin(a) + f cos(a)), where a = A / 180 radians.
c). Part c) shows the effect of a scaling of the axes by (3, 2). To make the figure clearer the new and old axes are
shown slightly displaced. Show that a point (e, f) in the new system lies at (3e, 2f) in the original system.
d). part d) shows a special case of scaling, a reflection about the x-axis. Show that the point (e, f) lies in the
original system at (e, -f).

5.5. Using Affine Transformations in a Program.

Chap 5. Transformations

9/28/99

page 35

We want to see how to apply the theory of affine transformations in a program to carry out scaling, rotating, and
translating of graphical objects. We also investigate how it is done when OpenGL is used. We look at 2D
examples first as they are easier to visualize, then move on to 3D examples.
To set the stage, suppose you have a routine house() that draws the house #1 in Figure 5.34. But you wish to
draw the version #2 shown that has been rotated through -300 and then translated through (32, 25). This is a
frequently encountered situation: an object is defined at a convenient size and position, but we want to draw it
(perhaps many times) at different sizes, orientations, and locations.
a).
y

23
#1
#2
x

Figure 5.34. Drawing a rotated and translated house.

As we discussed in Chapter 3, house() would draw the various polylines of the figure. If it were written in
raw OpenGL it might consist of a large number of chunks like:
glBegin(GL_LINES);
glVertex2d(V[0].x, V[0].y);
glVertex2d(V[1].x, V[1].y);
glVertex2d(V[2].x, V[2].y);
.... // the remaining points
glEnd();
based on some array V[] of points. Or if we use the Canvas class developed in Chapter 3 there would be a
number of calls to moveTo() and lineTo() as in (using the global canvas object cvs):
cvs.moveTo(V[0]);
cvs.lineTo(V[1]);
cvs.lineTo(V[2]);
... // the remaining points
In either case we would set up a world window and a viewport with calls like:
cvs.setWindow(...);
cvs.setViewport(...);
and we would be assured that all vertex positions V[i] are quietly converted from world coordinates to screen
window coordinates by the underlying window to viewport transformation.
But how do we arrange matters so that house #2 is drawn instead? There is the hard way and the easy way.
The hard way.
With this approach we construct the matrix for the desired transformation, say M, and build a routine, say
transform2D(), that transforms one point into another, such that:
Q =

transform2D(M, P);

Chap 5. Transformations

9/28/99

page 36

The routine produces Q = MP . To apply the transformation to each point V[i] in house() we must adjust the
source code above as in
cvs.moveTo(transform2D(M, V[0])); // move to the transformed point
cvs.lineTo(transform2D(M, V[1]));
cvs.lineTo(transform2D(M, V[2]));
...
so that the transformed points are sent to moveTo() and lineTo(). This is workable if the source code for
house() is at hand. But it is cumbersome at best, and not possible at all if the source code for house() is not
available. It also requires tools to create the matrix M in the first place.
The easy way.
We cause the desired transformation to be applied automatically to each vertex. Just as we know the window to
viewport mapping is quietly applied to each vertex as part of moveTo() and lineTo(), we can have an
additional transformation be quietly applied as well. It is often called the current transformation, CT. We
enhance moveTo() and lineTo() in the Canvas class so that they first quietly apply this transformation to the
argument vertex, and then apply the window to viewport mapping. (Clipping is performed at the world window
boundary as well.)
Figure 5.35 provides a slight elaboration of the graphics pipeline we introduced in Figure 5.7. When
glVertex2d()is called with argument V, the vertex V is first transformed by the CT to form point Q. Q is then
passed through the window to viewport mapping to form point S in the screen window. (As we see later, clipping
is also performed, inside this last mapping process.)

Figure 5.35. The current transformation is applied to vertices.

How do we extend moveTo() and lineTo() so they quietly carry out this additional mapping? (i.e. how do we
rewrite these functions in the Canvas class?) If you are not using OpenGL you must write code that actually
performs the transformation; this is described in Case Study 5.1. If you are using OpenGL it is done
automatically! OpenGL maintains a so-called modelview matrix, and every vertex that is passed down the
graphics pipeline is multiplied by it. We need only set up the modelview matrix to embody the desired
transformation.
OpenGL works entirely in 3D, so its modelview matrix produces 3D transformations. Here we work with the
modelview matrix in a restricted way to perform 2D transformations. Later we use its full power. Figure 5.36
shows how we restrict the 3D transformations to carry out the desired 2D transformations. The main idea is that
2D drawing is done in the xy-plane: the z-coordinate is understood to be zero. Therefore when we transform 2D
points we set the part of the underlying 3D transformation that affects the z-coordinate so that it has no effect at
all. For example, rotating about the origin in 2D is equivalent to rotating about the z-axis in 3D, as shown in the
figure. Further, although a scaling in 3D takes three scale factors, Sx, Sy, and Sz to scale in the x, y, and z
dimensions, respectively, we set the scale factor Sz = 1.

Chap 5. Transformations

9/28/99

page 37

Figure 5.36. 2D drawing takes place in the xy-plane.

The principal routines for altering the modelview matrix are glRotated()10, glScaled(), and
glTranslated(). These dont set the CT directly; instead each postmultiplies the CT (the modelview matrix)
by a particular matrix, say M, and puts the result back into the CT. That is, each of these routines creates a matrix
M as required for the new transformation, and performs:
CT = CT * M

(5.37)

The order is important. As we saw earlier, applying CT * M to a point is equivalent to first performing the
transformation embodied in M, followed by performing the transformation dictated by the previous value of CT.
Or if we are thinking in terms of transforming the coordinate system, it is equivalent to performing one
additional transformation to the existing current coordinate system.
OpenGL routines for applying transformations in the 2D case are:

glScaled(sx, sy, 1.0);

Postmultiply CT by a matrix that performs a scaling by sx in x and
by sy in y; Put the result back in CT. No scaling in z is done.

glTranslated(dx, dy, 0);

Postmultiply CT by a matrix that performs a translation by dx in x
and by dy in y; Put the result back in CT. No translation in z is done.

glRotated(angle, 0, 0, 1); Postmultiply CT by a matrix that performs a rotation through angle

degrees about the z-axis (indicated by (0, 0, 1))11. Put the result back in CT.

Since these routines only compose a transformation with the CT, we need some way to get started: to initialize
the CT to the identity transformation. OpenGL provides glLoadIdentity(). And because these functions can
be set to work on any of the matrices that OpenGL supports, we must inform OpenGL which matrix we are
altering. This is accomplished using glMatrixMode(GL_MODELVIEW).
Figure 5.37 shows suitable definitions of four new methods of the Canvas class that manage the CT and allow us
to build up arbitrarily complex 2D transformations. Their pleasing simplicity is possible because OpenGL is
doing the hard work.
//<<<<<<<<<<<<<<< initCT >>>>>>>>>>>>>>>>>
void Canvas:: initCT(void)
{
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
// set CT to the identity matrix
}
//<<<<<<<<<<<<<< scale2D >>>>>>>>>>>>>>>>>>>>
void Canvas:: scale2D(double sx, double sy)
{
glMatrixMode(GL_MODELVIEW);
glScaled(sx, sy, 1.0); // set CT to CT * (2D scaling)
}
//<<<<<<<<<<<<<<< translate2D >>>>>>>>>>>>>>>>>
void Canvas:: translate2D(double dx, double dy)
{
10 The suffix d indicates that its arguments are doubles. There is also the version glRotatef() that takes float arguments.
11 Here, as always, positive angles produce CCW rotations.

Chap 5. Transformations

9/28/99

page 38

glMatrixMode(GL_MODELVIEW);
glTranslated(dx, dy, 1.0); // set CT to CT * (2D translation)
}
//<<<<<<<<<<<<<<<< rotate2D >>>>>>>>>>>>>>>>>>>>
void Canvas:: rotate2D(double angle)
{
glMatrixMode(GL_MODELVIEW);
glRotated(angle, 0.0, 0.0, 1.0); // set CT to CT * (2D rotation)
}
Figure 5.37. Routines to manage the CT for 2D transformations.
We are now in a position to use 2D transformations. Returning to drawing version #2 of the house in Figure 5.34,
we next show the code that first rotates the house through -300 and then translates it through (32, 25). Notice that,
to get the ordering straight, it calls the operations in opposite order to the way they are applied: first the
translation operation, and then the rotation operation.
cvs.setWindow(...);
cvs.setViewport(..);
cvs.initCT();
house();
cvs.translate2D(32, 25);
cvs.rotate2D(-30.0);
house();

//
//
//
//
//
//

set the window to viewport mapping

get started with the identity transformation
draw the untransformed house first
CT now includes translation
CT now includes translation and rotation
draw the transformed house

Notice that we can scale, rotate, and position the house in any manner we choose, and never need to go inside
the routine house() or alter it. (In particular, the source code for house() need not be available.)
Some people find it more natural to think in terms of transforming the coordinate system. As shown in Figure
5.38 they would think of first translating the coordinate system through (32, 25) to form system #2, and then
rotating that system through -30o to obtain coordinate system #3. Because OpenGL applies transformations in the
order that coordinate systems are altered, the code for doing it this way first calls cvs.translate2D(32,
25) and then calls cvs.rotate2D(-30.0). This is, of course, identical to the code obtained doing it the
other way, but it has been arrived at through a different thinking process.

Figure 5.38. The same transformation viewed as a sequence of coordinate system changes.

We give some further examples to show how easily the CT is manipulated to produce various effects.
Example 5.5.1. Capitalizing on rotational symmetry.
Figure 5.39a shows a star made of stripes that seem to interlock with one another. This is easy to draw using
rotate2D(). Suppose that routine starMotif() draws a part of the star, the polygon shown in Figure 5.39b.
(Determining the positions of this polygons vertices is challenging, and is addressed in Case Study 5.2.) To
draw the whole star we just draw the motif five times, each time rotating the motif through an additional 72:

Chap 5. Transformations

9/28/99

page 39

a).

b).

(x1, y1 )

Figure 5.39. Using successive rotations of the coordinate system.

for(int count = 0; count < 5; count++)
{
starMotif();
cvs.rotate2D(72.0);
// concatenate another rotation
}
Visualize what is happening during each of these steps.
Example 5.5.2. Drawing snowflakes.
The beauty of a snowflake arises in good measure from its high degree of symmetry. A snowflake has six
identical spokes oriented 60o apart, and each spoke is symmetrical about its own axis. It is easy to produce a
complex snowflake by designing one half of a spoke, and drawing it 12 times. Figure 5.40a shows a snowflake,
based on the motif shown in Figure 5.40b. The motif is a polyline that meanders around above the positive xaxis. (To avoid any overlap with other parts of the snowflake, the polyline is kept below the 300 line shown in the
figure.) Suppose the routine flakeMotif() draws this polyline.
a).

b).

30 o line

Figure 5.40. Designing a Snowflake.

Each spoke of the snowflake is a combination of the motif and a reflected version. A reflection about the x-axis
is achieved by the use of scale2D(1,-1)(why?), so the motif plus its reflection can be drawn using
flakeMotif();
cvs.scale2D(1.0,-1.0);
flakeMotif();
cvs.scale2D(1.0,-1.0);

//
//
//
//

draw the top half

flip it vertically
draw the bottom half
restore the original axis

To draw the entire snowflake just do this six times, with an intervening rotation of 60o:
void drawFlake()
{
for(int count = 0; count < 6; count++) // draw a snowflake
{
flakeMotif();
cvs.scale2D(1.0,-1.0);
flakeMotif();

Chap 5. Transformations

c). after rotate2D()

d). after popCT()

Figure 5.43. Manipulating a stack of CTs.

The implementation of pushCT() and popCT() is simple when OpenGL is being used, since OpenGL has
routines glPushMatrix() and glPopMatrix() to manage several different stacks of matrices. Figure 5.44
shows the required functions. Note that each of them must inform OpenGL which matrix stack is being affected.
void Canvas:: pushCT(void)
{
glMatrixMode(GL_MODELVIEW);
glPushMatrix();
// push a copy of the top matrix
}
void Canvas:: popCT(void)
{
glMatrixMode(GL_MODELVIEW);
glPopMatrix();
// pop the top matrix from the stack
}
Figure 5.44. Routines to save and restore CTs.
Example 5.5.5: Tilings made easy.
Many beautiful designs called tilings appear on walls, pottery, and fabric. They are based on the repetition of a
basic motif both horizontally and vertically. Consider tiling the window with some motif, as suggested in Figure
5.45. The motif is drawn centered in its own coordinate system as shown in part a) using some routine
drawMotif(). Copies of the motif are drawn L units apart in the x-direction, and D units apart in the ydirection, as shown in part b).
a). the motif
b). the tiling

Figure 5.45. A tiling based on a motif.

Figure 5.46 shows how easily the coordinate system can be manipulated in a double loop to draw the tiling. The
CT is restored after drawing each row, so it returns to the start of that row, ready to move up to start the next row.
In addition, the whole block of code is surrounded with a pushCT() and a popCT(), so that after the tiling has
been drawn the CT is returned to its initial value, in case more drawing needs to be done.
cvs.pushCT();
// so we can return here
cvs.translate2D(W, H);
// position for the first motif
for(row = 0; row < 3; row++) // draw each row
{
cvs.pushCT();
for(col = 0; col < 4; col++)// draw the next row of motifs
{
motif();
cvs.translate2D(L, 0); // move to the right
}
cvs.popCT();
// back to the start of this row
cvs.translate2D(0, D);
// move up to the next row

Chap 5. Transformations

9/28/99

page 43

}
cvs.popCT();
Figure 5.46. Drawing a hexagonal tiling.

// back to where we started

Example 5.5.6. Using modeling transformations in a CAD program. Some programs must draw many
instances of a small collection of shapes. Figure 5.47 shows the example of a CAD program that analyzes the
behavior of an interconnection of digital logic gates. The user can construct a circuit by picking and placing
different gates at different places in the work area, possibly with different sizes and orientations. Each picture of
the object in the scene is called an instance of the object. A single definition of the object is given in a
coordinate system that is convenient for that object shape, called its master coordinate system. The
transformation that carries the object from its own master coordinate system to world coordinates to produce an
instance is often called a modeling transformation.

Create
Connect
Delete
Simulate
Quit
Work area
Figure 5.47. Creating instances in a pick-and-place application.
Figure 5.48 shows two logic gates, each defined once in its own master coordinate system. As the user creates
each instance of one of these gates, the appropriate modeling transformation is generated that orients and
positions the instance. The transformation might be stored simply as a set of parameters, say, S, A, dx, dy, with
the understanding that the modeling transformation would always consist of:
1.
2.
3.

A scaling by factor S;
A rotation through angle A
A translation through (dx, dy)

performed in that order. A list is kept of the gates in the circuit, along with the transformation parameters of each
gate.
a).

b).
y

NAND
2

NOT
x

1 2 3

Figure 5.48. Each gate type is defined in its own coordinate system.

Chap 5. Transformations

9/28/99

page 44

Whenever the drawing must be refreshed, each instance is drawn in turn, with the proper modeling
transformation applied. Code to do this might look like:
clear the screen
for(i = 0; i < numberOfGates; i++) // for each gate
{
pushCT(); // remember the CT
translate2D(dx[i], dy[i]); // apply the transformation
rotate2D(A[i]);
scale2D( S[i], S[i]);
drawGate(type[i]); // draw one of the two types
popCT(); // restore the CT
}
The CT is pushed before drawing each instance, so that it can be restored after the instance has been drawn. The
modeling transformation for the instance is determined by its parameters, and then one of the two gate shapes is
drawn. The necessary code has a simple organization, because the burden of sizing, orienting, and positioning
each instance has been passed to the underlying tools that maintain the CT and its stack.
Practice Exercises.
5.5.1. Developing the Transformations. Supposing OpenGL were not available, detail how you would write
the routines that perform elementary coordinate system changes:
void scale2D(double sx, double sy);
void translate2D(double dx, double dy);
5.5.2. Implementing the transformation stack. Define, in the absence of OpenGL, appropriate data types for a
stack of transformations, and write the routines pushCT() and popCT().
5.5.3. A hexagonal tiling. A hexagonal pattern provides a rich setting for tilings, since regular hexagons fit
together neatly as in a beehive. Figure 5.49 shows 9 columns of stacked 6-gons. Here the hexagons are shown
empty, but we could draw interesting figures inside them.
a). Show that the length of a hexagon with radius R is also R.
b). Show that the centers of adjacent hexagons in a column are separated vertically by 3 R and adjacent
columns are separated horizontally by 3 R / 2.
c). Develop code that draws this hexagonal tiling, using pushCT() and popCT() and suitable transformations to
keep track of where each row and column start.
y
R

x
Figure 5.49. A simple hexagonal tiling.

5.6. Drawing 3D Scenes with OpenGL

We introduced the drawing of 2D objects in Chapter 3, developing a simple class Canvas. This class provides
functions that establish a window and viewport, and that do line drawing through moveTo() and lineTo(). So
far in this chapter we have added the notion of the CT, and provided functions that perform 2D rotations,
scalings, and translations. These 2D transformations are really just special cases of 3D transformations: they
basically ignore the third dimension.

Chap 5. Transformations

9/28/99

page 45

In this section we examine how 3D transformations are used in an OpenGL-based program. The main emphasis
is on transforming objects in order to orient and position them as desired in a 3D scene. Not surprisingly it is all
done with matrices, and OpenGL provides the necessary functions to build the required matrices. Further, the
matrix stack maintained by OpenGL makes it easy to set up a transformation for one object, and then back up
to a previous transformation, in preparation for transforming another object.
It is very satisfying to build a program that draws different scenes using a collection of 3D transformations.
Experimenting with such a program also improves your ability to visualize what the various 3D transformations
do. OpenGL makes it easy to set up a camera that takes a snapshot of the scene from a particular point of
view. The camera is created with a matrix as well, and we study in detail in Chapter 7 how this is done. Here we
just use an OpenGL tool to set up a reasonable camera, so that attention can be focussed on transforming objects.
Granted we are using a tool before seeing exactly how it operates, but the payoff is high: you can make
impressive pictures of 3D scenes with a few simple calls to OpenGL functions.

5.6.1. An overview of the viewing process and the graphics pipeline.

All of our 2D drawing so far has actually used a special case of 3D viewing, based on a simple parallel
projection. We have been using the camera suggested in Figure 5.50. The eye that is viewing the scene
looks along the z-axis at the window, a rectangle lying in the xy-plane. The view volume of the camera is a
rectangular parallelepiped, whose four side walls are determined by the border of the window, and whose other
two walls are determined by a near plane and a far plane. Points lying inside the view volume are projected
onto the window along lines parallel to the z-axis. This is equivalent to simply ignoring their z-component, so
that the 3D point (x1 y1, z1) projects to (x1, y1, 0). Points lying outside the view volume are clipped off. A
separate viewport transformation maps the projected points from the window to the viewport on the display
device.

Figure 5.50. Simple viewing used in OpenGL for 2D drawing.

Now we move into 3D graphics and place 3D objects in a scene. For the examples here we continue to use a
parallel projection. (The more realistic perspective projection, for which more remote objects appear smaller than
nearby objects, is described in Chapter 7.) Therefore we use the same camera as in Figure 5.50, but allow it to
have a more general position and orientation in the 3D scene, in order to produce better views of the scene.
Figure 5.51 shows such a camera immersed in a scene, The scene consists of block, part of which lies outside the
view volume. The image produced by this camera is also shown.

Figure 5.51. A camera to produce parallel views of a scene.

We saw in the previous section that OpenGL provides the three functions glScaled(..),
glRotated(..), and glTranslated(..) for applying modeling transformations to a shape. The
block in Figure 5.51 is in fact a cube that has been stretched, rotated, and shifted as shown. OpenGL also
provides functions for defining the view volume and its position in the scene.
The graphics pipeline implemented by OpenGL does its major work through matrix transformations, so we first
give insight into what each of the matrices in the pipeline does. At this point it is important only to grasp the
basic idea of how each matrix operates: in Chapter 7 we give a detailed discussion. Figure 5.52 shows the
pipeline (slightly simplified). Each vertex of an object is passed through this pipeline with a call such as
glVertex3d(x, y, z). The vertex is multiplied by the various matrices shown, it is clipped if necessary,
and if it survives clipping it is ultimately mapped onto the viewport. Each vertex encounters three matrices:

Chap 5. Transformations

9/28/99

page 46

Figure 5.52. The OpenGL pipeline (slightly simplified).

The modelview matrix;

The projection matrix;
The viewport matrix;

The modelview matrix basically provides what we have been calling the CT. It combines two effects: the
sequence of modeling transformations applied to objects, and the transformation that orients and positions the
camera in space (hence its peculiar name modelview). Although it is a single matrix in the actual pipeline, it is
easier to think of it as the product of two matrices, a modeling matrix M, and a viewing matrix V. The modeling
matrix is applied first, and then the viewing matrix, so the modelview matrix is in fact the product VM (why?).
Figure 5.53 suggests what the M and V matrices do, for the situation introduced in Figure 5.51, where a camera
looks down on a scene consisting of a block. Part a shows a unit cube centered at the origin. A modeling
transformation based on M scales, rotates, and translates the cube into block shown in part b. Part b also shows
the relative position of the cameras view volume.
a). before model
b). after model
c) after model view
Figure 5.53. Effect of the modelview matrix in the graphics pipeline. a). Before the transformations. b). After the
modeling transformation. c). After the modelview transformation.
The V matrix is now used to rotate and translate the block into a new position. The specific transformation used
is that which would carry the camera from its position in the scene to its generic position, with the eye at the
origin and the view volume aligned with the z-axis, as shown in part c. The vertices of the block are now
positioned (that is, their coordinates have the proper values) so that projecting them onto a plane such as the near
plane yields the proper values for displaying the projected image. So the matrix V in fact effects a change of
coordinates of the scene vertices into the cameras coordinate system. (Camera coordinates are sometimes also
called eye coordinates. )
In the camera coordinate system the edges of the view volume are parallel to the x-, y-, and z-axes. The view volume
extends from left to right in x, from bottom to top in y, and from -near to -far in z, as shown. When the vertices of
the original cube have passed through the entire modelview matrix, they are located as shown in part c.
The projection matrix scales and shifts each vertex in a particular way, so that all those that lie inside the view
volume will lie inside a standard cube that extends from -1 to 1 in each dimension12. (When perspective
projections are being used this matrix does quite a bit more, as we see in Chapter 7.) This matrix effectively
squashes the view volume into the cube centered at the origin. This cube is a particularly efficient boundary
against which to clip objects, as we see in Chapter 7. Scaling the block in this fashion might badly distort it, of
course, but this distortion will be compensated for in the viewport transformation. The projection matrix also
reverses the sense of the z-axis, so that increasing values of z now represent increasing values of depth of a point
from the eye. Figure 5.54 shows how the block is transformed into a different block by this transformation.

Figure 5.54. Effect of the projection matrix (for parallel projections).

(Notice that the view volume of the camera need never be created as an object itself. It is defined only as that
particular shape that the projection matrix converts into the standard cube!)
Clipping is now performed, which eliminates the portion of the block that lies outside the standard cube.
Finally, the viewport matrix maps the surviving portion of the block into a 3D viewport. This matrix maps the
standard cube into a block shape whose x and y values extend across the viewport (in screen coordinates), and
whose z-component extends from 0 to 1 and retains a measure of the depth of point, as shown in Figure 5.55.
12 Coordinates in this system are sometimes called Normalized Device Coordinates.

Chap 5. Transformations

9/28/99

page 47

Figure 5.55. Effect of the viewport transformation.

This overview has outlined how the whole OpenGL graphics pipeline operates, showing that transformations are
a central ingredient. Each vertex of an object is subjected to a sequence of transformations that carry it from
world coordinates into eye coordinates, into a neutral system specially designed for clipping, and finally into the
right coordinate system for proper display. Each transformation is effected by a matrix multiplication.
Aside: Some important details of the pipeline have been suppressed in this first overview. When perspective
projections are used we need to include a perspective division that capitalizes on a property of homogeneous
coordinates. And of course many interesting details lurk in the last step of actually rendering the image, such as
computing the color of pixels in between vertices, and checking for proper hidden surface removal. These are
all addressed in later chapters.

5.6.2. Some OpenGL Tools for Modeling and Viewing.

We now see what functions OpenGL provides for modeling and setting the camera, and how to use them.
1). Three functions are used to set modeling transformations.
The following functions are normally used to modify the modelview matrix, so the modelview matrix is first
made current by executing: glMatrixMode(GL_MODELVIEW);

glScaled(sx, sy, sz); Postmultiply the current matrix by a matrix that performs a scaling by sx in
x, by sy in y, and by sz in z; Put the result back in the current matrix.

glTranslated(dx, dy, dz); Postmultiply the current matrix by a matrix that performs a
translation by dx in x, by dy in y, and by dz in z; Put the result back in the current matrix.

glRotated(angle, ux, uy, uz); Postmultiply the current matrix by a matrix that performs a
rotation through angle degrees about the axis that passes through the origin and the point (ux, uy, uz).13 Put
the result back in the current matrix. Equation 5.33 shows the matrix used to perform the rotation.

2). Setting the camera in OpenGL (for a parallel projection).

glOrtho(left,right,bott,top,near,far); Establishes as a view volume a parallelipiped that
extends from14 left to right in x, from bott to top in y, and from -near to -far in z. (Since this
definition operates in eye coordinates, the cameras eye is at the origin, looking down the negative z-axis.) This
function creates a matrix and postmultiplies the current matrix by it. (We show in Chapter 7 precisely what
values this matrix contains.)
Thus to set the projection matrix use:
glMatrixMode(GL_PROJECTION);
// make the projection matrix current
glLoadIdentity();
// set it to the identity matrix
glOrtho(left,right,bottom,top,near,far); // multiply it by the new matrix
Notice the minus signs above: because the default camera is located at the origin looking down the negative zaxis, using a value of 2 for near means to place the near plane at z = -2, that is, 2 units in front of the eye.
Similarly, using 20 for far places the far plane 20 units in front of the eye.
3). Positioning and aiming the camera.
13 Positive values of angle produce rotations that are CCW as one looks along the axis from the point (ux, uy, uz) towards the origin.
14 All parameters of glOrtho() are of type GLdouble.

Chap 5. Transformations

9/28/99

page 48

OpenGL offers a function that makes it easy to set up a basic camera:

gluLookAt(eye.x, eye.y, eye.z, look.x, look.y, look.z, up.x, up.y, up.z);
Creates the view matrix and postmultiplies the current matrix by it.
It takes as parameters the eye position, eye, of the camera and the look-at point, look. It also takes an
approximate up direction, up. Since the programmer knows where an interesting part of the scene is situated,
it is usually straightforward to choose reasonable values for eye and lookAt for a good first view. And up is
most often set to (0, 1, 0) to suggest an up direction parallel to the y-axis. Later we develop more powerful
tools for setting up a camera and for interactively flying it in an animation.
We want this function to set the V part of the modelview matrix VM. So it is invoked before any modeling
transformations are added, since subsequent modeling transformations will postmultiply the modelview matrix.
So to use gluLookAt() follow the sequence:
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
gluLookAt(eye.x, eye.y, eye.z,
look.x, look.y, look.z,
up.x, up.y, up.z)

// make the modelview matrix current

// start with a unit matrix
// the eye position
// the look at point
// the up direction

We discuss how this function operates in Chapter 7, and also develop more flexible tools for establishing the
camera. For those curious about what values gluLookAt() actually places in the modelview matrix, see the
exercises.
Example 5.6.1: Set up a typical camera. Cameras are often set to look down on the scene from some nearby
position. Figure 5.56 shows the camera with its eye situated at eye = (4,4,4) looking at the origin with lookAt
= (0,1,0). The up direction is set to up = (0, 1, 0). Suppose we also want the view volume to have a width of
6.4, a height of 4.8 (so its aspect ratio is 640/480), and to set near to 1 and far to 50. This camera would be
established using:

Figure 5.56. Setting a camera with gluLookAt().

glMatixMode(GL_PROJECTION); // set the view volume
glLoadIdentity();
glOrtho(-3.2, 3.2, -2.4, 2.4, 1, 50);
glMatrixMode(GL_MODELVIEW); // place and aim the camera
glLoadIdentity();
gluLookAt(2, 4, 5, 0, 0, 0, 0, 1, 0);
The exercises show the specific values that get placed in the modelview matrix.
Practice Exercises.
5.6.1. What does gluLookAt() do? We know that gluLookAt() builds a matrix that converts world
coordinates into eye coordinates. Figure 5.57 shows the camera as a coordinate system suspended in the world,
with its origin at eye, and oriented according to its three mutually perpendicular unit vectors u, v, and n. The eye
is looking in the direction -n. gluLookAt() uses the parameters eye, look, and up to create u, v, and n
according to

Figure 5.57*. Converting from world to camera coordinates.

Chap 5. Transformations

9/28/99

page 49

n = eye look
u = up n
v = nu

u
v
V=
n
0

(5.38)

and then normalizes all three of these to unit length. It builds the matrix
x
x
x

uy
vy
ny

uz
vz
nz

dx
dy
dz

where the point d has components ( d x , d y , d z ) = ( eye u, eye v, eye n) .

a). Show that u, v, and n are mutually perpendicular.
b). Show that matrix V properly converts world coordinates to eye coordinate by showing that it maps eye into
the origin (0,0,0,1)T, u into i = (1,0,0,0), v into j = (0,1,0), and n into k = (0,0,1).
c). Show for the case eye = (4, 4, 4), lookAt = (0,1,0), and up = (0,1,0) that the resulting matrix is:
.70711
0
0
.70711
.3313 .88345
.3313 .88345
V =

6.872
.4685
.6247
.6247

0
0
0
1

5.6.2. Inquiring of the values in a matrix in OpenGL. Print out the values in the modelview matrix to test
some of the assertions made about how it is formed. To see what is stored in the modelview matrix in OpenGL
define an array GLfloat mat[16] and use glGetFloatv(GL_MODELVIEW_MATRIX,mat) which
copies into mat[] the 16 values in the modelview matrix. M[i][j] is copied into the element mat[4j+i],
for i, j = 0, 1,..,3.

5.6.3. Drawing Elementary Shapes Provided by OpenGL

We see in the next chapter how to create our own 3D objects, but at this point we need some 3D objects to draw, in order to
experiment with setting and using cameras. The GLUT provides several ready-made 3D objects. These include a sphere, a
cone, a torus, the five Platonic solids (discussed in Chapter 6), and the famous teapot. Each is available as a wireframe
model and as a solid model with faces that can be shaded.

cube: glutWireCube(GLdouble size); Each side is of length size

sphere: glutWireSphere(GLdouble radius, GLint nSlices, GLint nStacks)
torus: glutWireTorus(GLdouble inRad, GLdouble outRad, GLint nSlices, GLint nStacks)
teapot: glutWireTeapot(GLdouble size)

There is also a glutSolidCube(), glutSolidSphere(), etc., that we use later. The shape of the torus is determined by i
inner radius inRad and outer radius outRad. The sphere and torus are approximated by polygonal faces, and you can adjus
parameters nSlices and nStacks to specify how many faces to use in the approximation. nSlices is the number of
subdivisions around the z-axis, and nStacks is the number of bands along the z-axis, as if the shape were a stack of nSta
disks.
Four of the Platonic solids (the fifth is the cube, already presented):
tetrahedron: glutWireTetrahedron()
octahedron:
glutWireOctahedron()
dodecahedron: glutWireDodecahedron()
icosahedron: glutWireIcosahedron()
All of the shapes above are centered at the origin.

cone:

glutWireCone(GLdouble baseRad, GLdouble height, GLint nSlices, GLint nStac

Chap 5. Transformations

9/28/99

page 50

tapered cylinder: gluCylinder(GLUquadricObj * qobj, GLdouble baseRad, GLdouble topRad,

GLdouble height, GLint nSlices, GLint nStacks)

The axes of the cone and tapered cylinder coincide with the z-axis. Their bases rest on the z = 0 plane, and they extend to z =
height along the z-axis. The radius of the cone and tapered cylinder at z = 0 is given by baseRad. The radius of the tapered
cylinder at z = height is topRad.
The tapered cylinder is actually a family of shapes, distinguished by the value of topRad. When topRad is 1 there is no ta
this is the classic cylinder. When topRad is 0 the tapered cylinder is identical to the cone.
Note that drawing the tapered cylinder in OpenGL requires some extra work, because it is a special case of a quadric surface,
we shall see in Chapter 6. To draw it you must 1) define a new quadric object, 2). set the drawing style (GLU_LINE for a
wireframe, GLU_FILL for a solid rendering), and 3). draw the object:
GLUquadricObj * qobj = gluNewQuadric();
// make a quadric object
gluQuadricDrawStyle(qobj,GLU_LINE);
// set style to wireframe
gluCylinder(qobj, baseRad, topRad, nSlices, nStacks); // draw the cylinder
Figure 5.58. Shapes available in the GLUT.
We next employ some of these shapes in two substantial examples that focus on using affine transformations to
model and view a 3D scene. The complete program to draw each scene is given. A great deal of insight can be
obtained if you enter these programs and produce the figures, and then see the effect of varying the various
parameters.
Example 5.6.2: A scene composed of wireframe objects.
Figure 5.59 shows a scene with several objects disposed at the corners of a unit cube. The cube has one corner at
the origin. Seven objects appear at various corners of the cube, all drawn as wireframes.
The camera is given a view volume that extends from -2 to 2 in y, with an aspect ratio of aspect = 640/480. Its
near plane is at N = 0.1, and its far plane is at F = 100. This is accomplished using:
glOrtho(-2.0* aspect, 2.0* aspect, -2.0, 2.0, 0.1, 100);
The camera is positioned with eye = (2, 2.2, 3), lookAt = (0, 0, 0), and up = (0, 1, 0) (parallel to the y-axis), using:
gluLookAt(2.0, 2.0, 2.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0);

Chap 5. Transformations

9/28/99

page 51

Figure 5.59. Wireframe drawing of various primitive shapes.

Figure 5.60 shows the complete program to produce the drawing. The main() routine initializes a 640 by 480 pixel screen
window, sets the viewport and background color, and specifies displayWire() as the display function to be called to
perform the drawing. In displayWire() the camera shape and position are established first. Then each object is drawn in
turn. Most objects need their own modeling matrix, in order to rotate and position them as desired. Before each modeling
transformation is established, a glPushMatrix() is used to member the current transformation, and after the object has been
drawn this prior current transformation is restored with a glPopMatrix(). Thus the code to draw each object is imbedded in
a glPushMatrix(), glPopMatrix() pair. Check each transformation carefully to see that it places its object at the proper
place.
Also shown in the figure are the x-, y-, and z- axes drawn with conical arrow heads. Displaying the underlying coordinate
system can help to orient the viewer. To draw the x-axis, the z-axis is rotated 90o about the y-axis to form a rotated system, and
the axis is redrawn in its new orientation. Note that this axis is drawn without immersing it in a glPushMatrix(),
glPopMatrix() pair, so the next rotation to produce the y-axis takes place in the already rotated coordinate system. Check
that its the proper rotation.
#include <windows.h> //suitable when using Windows 95/98/NT
#include <gl/Gl.h>
#include <gl/Glu.h>
#include <gl/glut.h>
//<<<<<<<<<<<<<<<<<<< axis >>>>>>>>>>>>>>
void axis(double length)
{ // draw a z-axis, with cone at end
glPushMatrix();
glBegin(GL_LINES);
glVertex3d(0, 0, 0); glVertex3d(0,0,length); // along the z-axis
glEnd();
glTranslated(0, 0,length -0.2);
glutWireCone(0.04, 0.2, 12, 9);
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<<<<<<<<<<<<< displayWire >>>>>>>>>>>>>>>>>>>>>>
void displayWire(void)
{
glMatrixMode(GL_PROJECTION); // set the view volume shape
glLoadIdentity();

Chap 5. Transformations

9/28/99

page 52

glOrtho(-2.064/48.0, 2.064/48.0, -2.0, 2.0, 0.1, 100);

glMatrixMode(GL_MODELVIEW); // position and aim the camera
glLoadIdentity();
gluLookAt(2.0, 2.0, 2.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0);
glClear(GL_COLOR_BUFFER_BIT); // clear the screen
glColor3d(0,0,0: // draw black lines
axis(0.5);
// z-axis
glPushMatrix();
glRotated(90, 0,1.0, 0);
axis(0.5);
// y-axis
glRotated(-90.0, 1, 0, 0);
axis(0.5);
// z-axis
glPopMatrix();
glPushMatrix();
glTranslated(0.5, 0.5, 0.5); // big cube at (0.5, 0.5, 0.5)
glutWireCube(1.0);
glPopMatrix();
glPushMatrix();
glTranslated(1.0,1.0,0);
glutWireSphere(0.25, 10, 8);
glPopMatrix();

// sphere at (1,1,0)

glPushMatrix();
glTranslated(1.0,0,1.0);
// cone at (1,0,1)
glutWireCone(0.2, 0.5, 10, 8);
glPopMatrix();
glPushMatrix();
glTranslated(1,1,1);
glutWireTeapot(0.2); // teapot at (1,1,1)
glPopMatrix();
glPushMatrix();
glTranslated(0, 1.0 ,0); // torus at (0,1,0)
glRotated(90.0, 1,0,0);
glutWireTorus(0.1, 0.3, 10,10);
glPopMatrix();
glPushMatrix();
glTranslated(1.0, 0 ,0); // dodecahedron at (1,0,0)
glScaled(0.15, 0.15, 0.15);
glutWireDodecahedron();
glPopMatrix();
glPushMatrix();
glTranslated(0, 1.0 ,1.0); // small cube at (0,1,1)
glutWireCube(0.25);
glPopMatrix();
glPushMatrix();
glTranslated(0, 0 ,1.0); // cylinder at (0,0,1)
GLUquadricObj * qobj;
qobj = gluNewQuadric();
gluQuadricDrawStyle(qobj,GLU_LINE);
gluCylinder(qobj, 0.2, 0.2, 0.4, 8,8);
glPopMatrix();
glFlush();
}
//<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char **argv)
{

Chap 5. Transformations

9/28/99

page 53

glutInit(&argc, argv);
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB );
glutInitWindowSize(640,480);
glutInitWindowPosition(100, 100);
glutCreateWindow("Transformation testbed - wireframes");
glutDisplayFunc(displayWire);
glClearColor(1.0f, 1.0f, 1.0f,0.0f); // background is white
glViewport(0, 0, 640, 480);
glutMainLoop();
}
Figure 5.60. Complete program to draw Figure 5.65 using OpenGL.
Notice that the sides of the large cube that are parallel in 3D are also displayed as parallel. This is a result of using a parallel
projection. The cube looks slightly unnatural because we are used to seeing the world with a perspective projection. As we see
in Chapter 7, if a perspective projection were used instead, these parallel edges would not be drawn parallel.
Example 5.6.3. A 3D scene rendered with shading.
We develop a somewhat more complex scene to illustrate further the use of modeling transformations. We also show how eas
OpenGL makes it to draw much more realistic drawings of solid objects by incorporating shading, along with proper hidden
surface removal.
Two views of a scene are shown in Figure 5.61. Both views use a camera set by gluLookAt(2.3, 1.3, 2,
0, 0.25, 0, 0.0,1.0,0.0). Part a uses a large view volume that encompasses the whole scene; part b
uses a small view volume that encompasses only a small portion of the scene, thereby providing a close-up view.
The scene contains three objects resting on a table in the corner of a room. Each of the three walls is made by flattening a
cube into a thin sheet, and moving it into position. (Again, they look somewhat unnatural due to the use of a parallel
projection.) The jack is composed of three stretched spheres oriented at right angles plus six small spheres at their ends.

Figure 5.61. A simple 3D scene - a). Using a large view volume. b). Using a small view volume.
The table consists of a table top and four legs. Each of the tables five pieces is a cube that has been scaled to the desired size
and shape. The layout for the table is shown in Figure 5.62. It is based on four parameters that characterize the size of its parts
topWidth, topThick, legLen, and legThick. A routine tableLeg() draws each leg, and is called four times within
the routine table() to draw the legs in the four different locations. The different parameters used produce different modeling
transformations within tableLeg(). As always, a glPushMatrix(), glPopMatrix() pair surrounds the modeling
functions to isolate their effect.
a).
b).

Chap 5. Transformations

9/28/99

page 54

Figure 5.62. Designing the table.

The complete code for this program is shown in Figure 5.63. Note that the solid version of each shape, such as
glutSolidSphere(), is used here, rather than the wire version. Check the assorted transformations used to orient and
position each object in the scene. Check the jack model particularly. This example is designed to whet your appetite for trying
out scenes on your own, practicing with transformations.
The code also shows the various things that must be done to create shaded images. The position and properties of a light
source must be specified, along with certain properties of the objects surfaces, in order to describe how they reflect light.
Since we discuss the shading process in Chapter 8 we just present the various function calls here; using them as shown will
generate shading.
#include <windows.h>
#include <iostream.h>
#include <gl/Gl.h>
#include <gl/Glu.h>
#include <gl/glut.h>
//<<<<<<<<<<<<<<< wall >>>>>>>>>>>>>>>>
void wall(double thickness)
{ // draw thin wall with top = xz-plane, corner at origin
glPushMatrix();
glTranslated(0.5, 0.5 * thickness, 0.5);
glScaled(1.0, thickness, 1.0);
glutSolidCube(1.0);
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<< tableLeg >>>>>>>>>>>>>>>>>>>
void tableLeg(double thick, double len)
{
glPushMatrix();
glTranslated(0, len/2, 0);
glScaled(thick, len, thick);
glutSolidCube(1.0);
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<<<<< jack part >>>>>>>>>>>>>
void jackPart()
{ // draw one axis of the unit jack - a stretched sphere
glPushMatrix();
glScaled(0.2,0.2,1.0);
glutSolidSphere(1,15,15);
glPopMatrix();
glPushMatrix();
glTranslated(0,0,1.2); // ball on one end
glutSolidSphere(0.2,15,15);
glTranslated(0,0, -2.4);
glutSolidSphere(0.2,15,15); // ball on the other end
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<<< jack >>>>>>>>>>>>>>>>>>>>
void jack()
{ // draw a unit jack out of spheroids
glPushMatrix();
jackPart();
glRotated(90.0, 0, 1, 0);
jackPart();
glRotated(90.0, 1,0,0);
jackPart();
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<<<<<<< table >>>>>>>>>>>>>>>>>>>>

Chap 5. Transformations

9/28/99

page 55

void table(double topWid, double topThick, double legThick, double legLen)

{ // draw the table - a top and four legs
glPushMatrix(); // draw the table top
glTranslated(0, legLen, 0);
glScaled(topWid, topThick, topWid);
glutSolidCube(1.0);
glPopMatrix();
double dist = 0.95 * topWid/2.0 - legThick / 2.0;
glPushMatrix();
glTranslated(dist, 0, dist);
tableLeg(legThick, legLen);
glTranslated(0, 0, -2 * dist);
tableLeg(legThick, legLen);
glTranslated(-2 * dist, 0, 2*dist);
tableLeg(legThick, legLen);
glTranslated(0, 0, -2*dist);
tableLeg(legThick, legLen);
glPopMatrix();
}
//<<<<<<<<<<<<<<<<<<<<< displaySolid >>>>>>>>>>>>>>>>>>>>>>
void displaySolid(void)
{
//set properties of the surface material
GLfloat mat_ambient[] = { 0.7f, 0.7f, 0.7f, 1.0f}; // gray
GLfloat mat_diffuse[] = {0.6f, 0.6f, 0.6f, 1.0f};
GLfloat mat_specular[] = {1.0f, 1.0f, 1.0f, 1.0f};
GLfloat mat_shininess[] = {50.0f};
glMaterialfv(GL_FRONT,GL_AMBIENT,mat_ambient);
glMaterialfv(GL_FRONT,GL_DIFFUSE,mat_diffuse);
glMaterialfv(GL_FRONT,GL_SPECULAR,mat_specular);
glMaterialfv(GL_FRONT,GL_SHININESS,mat_shininess);
// set the light source properties
GLfloat lightIntensity[] = {0.7f, 0.7f, 0.7f, 1.0f};
GLfloat light_position[] = {2.0f, 6.0f, 3.0f, 0.0f};
glLightfv(GL_LIGHT0, GL_POSITION, light_position);
glLightfv(GL_LIGHT0, GL_DIFFUSE, lightIntensity);
// set the camera
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
double winHt = 1.0; // half-height of the window
glOrtho(-winHt*64/48.0, winHt*64/48.0, -winHt, winHt, 0.1, 100.0);
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
gluLookAt(2.3, 1.3, 2, 0, 0.25, 0, 0.0,1.0,0.0);
// start drawing
glClear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT); // clear the screen
glPushMatrix();
glTranslated(0.4, 0.4, 0.6);
glRotated(45,0,0,1);
glScaled(0.08, 0.08, 0.08);
jack();
// draw the jack
glPopMatrix();
glPushMatrix();
glTranslated(0.6, 0.38, 0.5);
glRotated(30,0,1,0);
glutSolidTeapot(0.08);
// draw the teapot
glPopMatrix();
glPushMatrix();
glTranslated(0.25, 0.42, 0.35);// draw the sphere
glutSolidSphere(0.1, 15, 15);
glPopMatrix();
glPushMatrix();
glTranslated(0.4, 0, 0.4);

Chap 5. Transformations

9/28/99

page 56

table(0.6, 0.02, 0.02, 0.3); //

glPopMatrix();
wall(0.02);
//
glPushMatrix();
glRotated(90.0, 0.0, 0.0, 1.0);
wall(0.02);
//
glPopMatrix();
glPushMatrix();
glRotated(-90.0,1.0, 0.0, 0.0);
wall(0.02);
//
glPopMatrix();
glFlush();

draw the table

wall #1: in xz-plane
wall #2: in yz-plane

wall #3: in xy-plane

}
//<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char **argv)
{
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_SINGLE | GLUT_RGB| GLUT_DEPTH);
glutInitWindowSize(640,480);
glutInitWindowPosition(100, 100);
glutCreateWindow("shaded example - 3D scene");
glutDisplayFunc(displaySolid);
glEnable(GL_LIGHTING); // enable the light source
glEnable(GL_LIGHT0);
glShadeModel(GL_SMOOTH);
glEnable(GL_DEPTH_TEST); // for hidden surface removal
glEnable(GL_NORMALIZE); // normalize vectors for proper shading
glClearColor(0.1f,0.1f,0.1f,0.0f); // background is light gray
glViewport(0, 0, 640, 480);
glutMainLoop();
}
Figure 5.63. Complete program to draw the shaded scene.
Practice Exercises.
5.6.3. Inquiring of the values in a matrix in OpenGL. Test some of the assertions above that put specific
values into the modelview matrix. You can see what is stored in the modelview matrix in OpenGL by defining an
array GLfloat mat[16] and using glGetFloatv(GL_MODELVIEW_MATRIX,mat) which copies into
mat[] the 16 values in the modelview matrix. M[i][j] is copied into the element mat[4j+i], for i, j = 0,
1,..,3.

5.6.4. Reading a Scene Description from a File.

In the previous examples the scene was described through specific OpenGL calls that transform and draw each
object, as in
glTranslated(0.25, 0.42, 0.35);
glutSolidSphere(0.1, 15, 15); // draw a sphere
The objects in the scene were therefore hard-wired into the program. This way of specifying a scene is
cumbersome and error-prone. It is a boon when the designer can specify the objects in a scene through a simple
language, and place the description in a file. The drawing program becomes a (much simpler) general-purpose
program: it reads a scene file at run-time and draws whatever objects are encountered in the file.
The Scene Description Language (SDL) described in Appendix 4 provides such a tool. We define a Scene class
also described in Appendix 4, and available on the books web site - that supports the reading of an SDL file
and the drawing of the objects described in the file. It is very simple to use the Scene class in an application: a
global Scene object is created:
Scene scn; // create a scene object

Chap 5. Transformations

9/28/99

page 57

and the read() method of the class is called to read in a scene file:
scn.read("simple.dat");

// read the scene file & build an object list

Figure 5.64 shows the data structure for the scn object, after the following simple SDL file has been read.

Figure 5.64. An object of the Scene class.

! simple.dat: a simple scene having one light and four shapes
background 0 0 1
! give the scene a blue background
light 2 9 8 1 1 1
! put a white light at (9, 9, 9)
diffuse .9 .1 .1
! make the following objects reddish
translate 3 5 2 sphere
! put a sphere at 3 5 2
translate 4 6 8 cone ! put a cone in the scene
translate 1 1 1 cube
! add a cube
diffuse 0 1 0
! make the following objects green
translate 40 5 2 scale .2 .2 .2 sphere !add a tiny sphere
The first line is a comment; comments extend to the end of the line. This scene has a bright blue background
color (red, green, blue) = (0, 0, 1), a bright white (1, 1, 1) light situated at (2, 9, 8), and four objects: two spheres,
a cone and a cube. The light field points to the list of light sources, and the obj field points to the object list.
Each shape object has its own affine transformation M that describes how it is scaled, rotated, and positioned in
the scene. It also contains various data fields that specify its material properties, that are important when we want
to render it faithfully as discussed in Chapter 8. Only the diffuse field is shown in the figure.
Once the light list and object list have been built, the application can render the scene:
scn.makeLightsOpenGL(),
scn.drawSceneOpenGL(); // render the scene using OpenGL
The first instruction passes a description of the light sources to OpenGL. The second uses the method
drawSceneOpenGL() to draw each object in the object list. The code for this method is very simple:
void Scene :: drawSceneOpenGL()
{
for(GeomObj* p = obj; p ; p = p->next)
p->drawOpenGL(); // draw it
}
It moves a pointer through the object list, calling drawOpenGL() for each object in turn. Its a nice example of
using polymorphism which is a foundation-stone of object-oriented programming: Each different shape knows
how to draw itself: it has a method drawOpenGL() that calls the appropriate routine for that shape. So when p
points to a sphere the drawOpenGL() routine for a sphere is called automatically; when p points to a cone the
drawOpenGL() routines for a cone is called, etc. Figure 5.65 shows the methods for the Sphere and Cone
classes; they differ only in the final OpenGL drawing routine that is called. Each first passes the objects material
properties to OpenGL, then updates the modelview matrix with the objects specific affine transformation. The
original modelview matrix is pushed and later restored, to protect it from being affected after this object has been
drawn.
void Sphere :: drawOpenGL()
{
tellMaterialsGL(); //pass material data to OpenGL
glPushMatrix();
glMultMatrixf(transf.m); // load this objects matrix
glutSolidSphere(1.0,10,12); // draw a sphere
glPopMatrix();
}
void Cone :: drawOpenGL()

Chap 5. Transformations

9/28/99

page 58

{
tellMaterialsGL();//pass material data to OpenGL
glPushMatrix();
glMultMatrixf(transf.m); // load this objects matrix
glutSolidCone(1.0,1.0, 10,12); // draw a cone
glPopMatrix();
}
Figure 5.65. the drawOpenGL() methods for two shapes.
Figure 5.66 shows the program that reads an SDL file and draws the scene. It is very short, (but of course the
code for the classes Scene, Shape, etc. must be loaded as well). It reads the particular SDL file
myScene1.dat, which recreates the same scene as in Figure 5.63. Note that by simply changing the SDL file
that is read this program can draw any scene described in SDL, without any changes in its code.
#include "basicStuff.h"
#include "Scene.h"
//############################ GLOBALS ########################
Scene scn; // construct the scene object
//<<<<<<<<<<<<<<<<<<<<<<< displaySDL >>>>>>>>>>>>>>>>>>>>>>>>>>
void displaySDL(void)
{
glMatrixMode(GL_PROJECTION); //set the camera
glLoadIdentity();
double winHt = 1.0; // half-height of the window
glOrtho(-winHt*64/48.0, winHt*64/48.0, -winHt, winHt, 0.1, 100.0);
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
gluLookAt(2.3, 1.3, 2, 0, 0.25, 0, 0.0,1.0,0.0);
glClear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT);
// clear screen
scn.drawSceneOpenGL();
} // end of display
//<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char **argv)
{
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_RGB |GLUT_DEPTH);
glutInitWindowSize(640, 480);
glutInitWindowPosition(100, 100);
glutCreateWindow("read and draw an SDL scene");
glutDisplayFunc(displaySDL);
glShadeModel(GL_SMOOTH);
glEnable(GL_DEPTH_TEST);
glEnable(GL_NORMALIZE);
glViewport(0, 0, 640, 480);
scn.read("myScene1.dat"); //read the SDL file and build the objects
glEnable(GL_LIGHTING);
scn.makeLightsOpenGL(); // scan the light list and make OpenGL lights
glutMainLoop();
}
Figure 5.66. Drawing a scene read in from an SDL file.
The SDL file that describes the scene of Figure 5.63 is shown in Figure 5.67. It defines the jack shape of nine
spheres by first defining a jackPart and then using it three times, as explained in Appendix 4. Similarly a leg
of the table is first defined as a unit, and then used four times.
! myScene1.dat
light 20 60 30 .7 .7 .7 !put a light at (20,60,30),color:(.7, .7, .7)
ambient .7 .7 .7 ! set material properties for all of the objects
diffuse .6 .6 .6
specular 1 1 1
exponent 50
def jackPart{ push scale .2 .2 1 sphere pop

Chap 5. Transformations

9/28/99

page 59

push translate 0 0 1.2 scale .2 .2 .2 sphere pop

push translate 0 0 -1.2 scale .2 .2 .2 sphere pop
}
def jack{ push use jackPart
rotate 90 0 1 0 use jackPart
rotate 90 1 0 0 use jackPart pop
}
def wall{push translate 1 .01 1 scale 1 .02 1 cube pop}
def leg {push translate 0 .15 0 scale .01 .15 .01 cube pop}
def table{
push translate 0 .3 0 scale .3 .01 .3 cube pop !table top
push
translate .275 0 .275 use leg
translate 0 0 -.55 use leg
translate -.55 0 .55 use leg
translate 0 0 -.55 use leg pop
}
!now add the objects themselves
push translate .4 .4 .6 rotate 45 0 0 1 scale .08 .08 .08 use jack pop
push translate .25 .42 .35 scale .1 .1 .1 sphere pop
push translate .6 .38 .5 rotate 30 0 1 0 scale .08 .08 .08 teapot pop
push translate 0.4 0 0.4 use table pop
use wall
push rotate 90 0 0 1 use wall pop
push rotate -90 1 0 0 use wall pop
Figure 5.67. The SDL file to create the scene of Figure 5.63.
With a scene description language like SDL available, along with tools to read and parse it, the scene designer
can focus on creating complex scenes without having to work at the application code level. The scene being
developed can be edited and tested again and again until it is right. The code developer puts the initial effort into
constructing an application that can render any scene describable in SDL.

5.7. Summary of the Chapter.

Affine transformations are a staple in computer graphics, for they offer a unified tool for manipulating graphical
objects in important ways. A designer always needs to scale, orient, and position objects in order to compose a
scene as well as an appropriate view of the scene, and affine transformations make this simple to manage in a
program.
Affine transformations convert one coordinate frame into another, and when homogeneous coordinates are used,
an affine transformation is captured in a single matrix form. A sequence of such transformations can be
combined into a single transformation whose matrix is simply the product of the individual transformation
matrices. Significantly, affine transformations preserve straightness, so the image of a line is another line, and
the image of a plane is a plane. This vastly simplifies working with lines and planes in a program, where one
benefits from the simple representations of a line (its two endpoints) and a plane (by three points or four
coefficients). In addition, parallelism is preserved, so that parallelograms map to parallelograms, and in 3D
parallelepipeds map to parallelepipeds. This makes it simpler to visualize the geometric effects of affine
transformations.
Three dimensional affine transformations are much more complex than their 2D counterparts, particularly when
it comes to visualize a combination of rotations. A given rotation can be viewed as three elementary rotations
through Euler angles, or a rotation about some axis, or simply as a matrix that has special properties. (Its columns
are orthogonal unit vectors). It is often important to move between these different forms.
OpenGL and other graphics packages offer powerful tools for manipulating and applying transformations. In
OpenGL all points are passed through several transformations, and the programmer can exploit this to define and

Chap 5. Transformations

9/28/99

page 60

manipulate a camera, as well as to size and position different objects into a scene. Two of the matrices used by
OpenGL (the modelview and viewport transformations) define affine transformations, whereas the projection
matrix normally defines a perspective transformation, to be examined thoroughly in Chapter 7. OpenGL also
maintains a stack of transformations, which make it easy for the scene designer to control the dependency of one
objects position on that of another, and to create objects that are composed of several related parts.
The SDL language, along with the Scene and Shape classes, make it much simpler to separate programming
issues from scene design issues. An application is developed once that can draw any scene described by a list of
light sources and a list of geometric objects. This application is then used over and over again with different
scene files. A key task in the scene design process is applying the proper geometric transformations to each
object. Since a certain amount of trial and error is usually required, it is convenient to be able to express these
transformations in a concise and readable way.
The next section presents a number of Case Studies that elaborate on the main ideas of the chapter and suggest
ways to practice with affine transformations in a graphics program. These range from plunging deeper into the
theory of transformations to actual modeling and rendering of objects such as electronic CAD circuits and robots.

5.8 Case Studies.

Case study 5.1. Doing your own transforming by the CT in Canvas.
(Level of Effort: II). Its easy to envision situations where you must implement the transformation mechanism
yourself, rather than rely on OpenGL to do it. In this Case Study you add the support of a current transformation
to the Canvas class for 2D drawing. This involves writing several functions, those to initialize and alter the CT
itself:
void
void
void
void

Canvas::
Canvas::
Canvas::
Canvas::

initCT(void);
// init CT to unit transformation
scale2D(double sx, double sy);
translate2D(double dx, double dy);
rotate2D(double angle);

as well as others that are incorporated into moveTo() and lineTo() so that all points sent to them are silently
transformed before being used. For extra benefit add the stack mechanism for the CT as well, along with functions
pushCT() and popCT(). Exercise your new tools on some interesting 2D modeling and drawing examples.

Case Study 5.2. Draw the star of Fig 5.39. using multiple rotations.
(Level of Effort: I). Develop a function that draws the polygon in figure 5.39b that is one fifth of the star. Use it
with rotation transformations to draw the whole star.

Case Study 5.3. Decomposing a 2D Affine Transformation.

(Level of Effort: II). We have seen that some affine transformations can be expressed as combinations of others.
Here we decompose any 2D affine transformation into a combination of scalings, rotations, and translations.
We also show that a rotation can be performed by three successive shears, which gives rise to a very fast arc
drawing routine. Some results on what lurks in any 3D affine transformation are also discussed. We just
summarize results here.
Because an affine transformation is a linear transformation followed by an offset we omit the translation part of
the affine transformation, and focus on what a linear transformation is. So we use 2 by 2 matrices here for
simplicity.

a). Two 2D linear transformations.

Consider the 2 by 2 matrix M that represents a 2D linear transformation. Matrix M can always be factored into a
rotation, a scaling , and a shear. Call the four scalars in M by the names a, b, c, and d for brevity. Verify by direct
multiplication that M is the product of the three matrices [martin82]:

Chap 5. Transformations

9/28/99

page 61

a
c

b 1
= ac + bd
d R2

0 R
1 0

a
0
ad bc R
b
R
R

b
R
a
R

(5.39)

where R = a2 + b2 . The leftmost matrix on the right hand side is recognized as a shear, the middle one as a
scaling, and the rightmost one as a rotation (why?).
Thus any 2D affine transformation is a rotation followed by a scaling followed by a shear followed by a
translation.
An alternative decomposition, based on the so-called Gram-Schmidt process, is discussed in Case Study 5.5.

4 3
into the product of shear, scaling, and rotation matrices.
2 7

Example 5.8.1. Decompose the matrix M =

Solution: Check the following, obtained by direct substitution in Equation 5.39:

0 5
0 4 / 5 3 / 5
4 3 1

M=
=
2 7 13 / 25 1 0 34 / 5 3 / 5 4 / 5
b). A 2D Rotation is three shears.
Matrices can be factored in different ways. In fact a rotation matrix can be written as the product of three shear
matrices! [Paeth90] This leads to a particularly fast method for performing a series of rotations, as we will see.
Equation 5.40 shows a rotation represented as three successive shears. It can be verified by direct multiplication.
It demonstrates that a rotation is a shear in y followed by a shear in x, followed by a repetition of the first shear.

cos(a) sin(a)
-sin(a) cos(a)

1
0

tan(a/2)
1

1
-sin(a)

0
1

1
0

tan(a/2)
1

(5.40)

Calling T = tan(a/2), and S = sin (a) show that we can write the sequence of operations that rotate point (x, y)
as15:
x = T * y + x;
y = y;

{first shear}

x = x;
y = y - S *x;

{second shear}

x = T * y + x;
y = y;

{third shear}

using primes to distinguish new values from old. But operations like x = x do nothing, so the primes are
unnecessary and this sequence reduces to:
x = x + T * y;
y = y - S * x;
x = x + T *y;

{actual operations for the three shears}

15Note that sin(a) may be found quickly from tan(a/2) by one multiplication and one division (see Appendix 2): S = (2T)/(1 +

T2)

Chap 5. Transformations

9/28/99

page 62

If we have only one rotation to perform there is no advantage to going about it this way. However, it becomes
very efficient if we need to do a succession of rotations through the same angle. Two places where we need to do
this are 1). calculating the vertices of an n-gon which are equispaced points around a circle, and 2). computing
the points along the arc of a circle.
Figure 5.68 shows a code fragment for calculating the positions of n points around a circle. It loads the
successive values (cos(i * 2/n + b), sin(i * 2/n + b)) into the array of points p[].
T = tan(PI/n);
// tangent of half angle16
S = 2 * T/(1 + T * T);
// sine of angle
p[0].x = sin(b);
// initial angle, vertex 0
p[0].y = cos(b);
for (int i = 1; i < n; i++)
{
p[i].y = p[i-1].x * T + p[i-1].y; //1st shear
p[i].x = p[i-1].x -S * p[i-1].y; //2nd shear
p[i].y = p[i-1].x * T + p[i-1].y; //3rd shear
}
Figure 5.68. Building the vertices of an n-gon efficiently.
Figure 5.69 shows how to use shears to build a fast arc drawer. It does the same job as drawArc() in Chapter
3, but much more efficiently, since it avoids the repetitive computation of sin() and cos().
void drawArc2(RealPoint c, double R,
double startangle, double sweep) // in degrees
{
#define n 30
#define RadPerDeg .01745329
double delang = RadPerDeg * sweep / n;
double T = tan(delang/2);
// tan. of half angle
double S = 2 * T/(1 + T * T);
// sine of half angle
double snR = R * sin(RadPerDeg * startangle);
double csR = R * cos(RadPerDeg * startangle);
moveTo(c.x + csR, c.y + snR);
for(int i = 1; i < n; i++)
{
snR += T * csR;
// build next snR, csR pair
csR -= S * snR;
snR += T * csR;
lineTo(c.x + csR, c.y + snR);
}
}
Figure 5.69. A fast arc drawer.
Develop a test program that uses this routine to draw arcs. Compare its efficiency with arc drawers that compute
each vertex using trigonometry.

c). Is a shear fundamental?

One sometimes reads in the graphics literature that the fundamental elementary transformations are the rotation,
scaling, and translation; shears seem to be second class citizens. This attitude may stem from the fact that any
shear can be decomposed into a combination of rotations and scalings. The following equation may be verified
by multiplying out the three matrices on the right [Greene90]:

a - 1a

1
1 + a2

1 -a
a 1

a
0

0
1/a

a 1
-1 a

1
1 + a2
(5.41)

Note: Some C environments do not support tan() directly. Use sin()/cos() - of course first checking that this denominator is not zero.

Chap 5. Transformations

9/28/99

page 63

The middle matrix is a scaling, while the outer two matrices (when combined with the scale factors shown) are
rotations. For the left-hand rotation associate 1/ 1 + a2 with cos() and -a / 1 + a2 with sin() for some
angle . Thus tan() = -a. Similarly for the right-hand rotation associate cos() and sin) for angle , where
tan() = 1/a. Note that and are related: = + / 2 (why?).
Using this decomposition, write the shear in Equation 5.41 as a rotation then a scaling then a rotation to
conclude that every 2D affine transformation is the following sequence of elementary transformations:
Any affine transformation = Scale * Rotation * Scale * Rotation * Translation

(5.42)

Practice Exercises
5.8.1. A golden decomposition. Consider the special case of a unit shear where the term a - 1/a in Equation
5.3.5 is 1. What value must a have? Determine the two angles and associated with the rotations.
Solution: Since a must satisfy a = 1 + 1/a, a is the golden ratio !. Thus

1 0 = cos( ) -sin()
sin( ) cos( )
1 1

0 1

cos( ) sin( )
-sin( ) cos( )

(5.43)
where = tan-1() = 58.28 and = tan - 1(1/) = 31.72.
5.8.2. Unit Shears. Show that any shear in x contains within it a unit shear. Decompose the shear given by

1 0
h 1

into the product of a scaling, a unit shear, and another scaling.

5.8.3. Seeing It Graphically. Drawing upon the ideas in Exercise 5.8.1, draw a rectangle on graph paper; rotate
o
o
it through - 58.28 ; scale it by (, 1/); and finally rotate it through 31.72 . Sketch each intermediate result, and
show that the final result is the same parallelogram obtained when the original rectangle undergoes a unit shear.
5.8.4. Decompose a Transformation. Decompose the transformation
Qx = 3 Px - 2 Py + 5
Qy = 4 Px + Py - 6
into a product of rotations, scalings, and translations.
5.8.5. One reflection can always be in the x-axis. Show that a rotation through angle A about the origin may
always be effected by a reflection in the x-axis followed by a reflection in a line at angle A/2.
5.8.6. Isometries. A particularly important family of affine transformations in the study of symmetry are the
isometries (same measure), since they dont alter the distance between two points and their images. For any
two points P and Q the distance |T(P) - T(Q)| is the same as the distance |P - Q| when T() is an isometry. Show
that if T() is affine with matrix M , then T is an isometry if and only if each row of M considered as a vector is of
unit length, and the two rows are orthogonal.
Solution: Call r = P - Q. Then |T(P) - T(Q)| = |PM + d - QM - d| = |rM| = |(rxm11 + rym21, rxm12 + rym22)|.
Equating |rM|2 to |r|2 requires m 2 + m 2 = 1, m 2 + m 2 = 1, and m m + m m = 0, as claimed.
11

5.8.7. Ellipses Are Invariant. Show that ellipses are invariant under an affine transformation. That is, if E is an
ellipse and if T is an affine transformation, then the image T(E) of the points in E also makes an ellipse. Hint to
Solution: Any affine is a combination of rotations and scalings. When an ellipse is rotated it is still an ellipse, so
only the non-uniform scalings could possibly destroy the ellipse property. So it is necessary only to show that an
ellipse, when subjected to a non-uniform scaling, is still an ellipse.
5.8.8. What else is invariant? Consider what class of shapes (perhaps a broader class than ellipses) is invariant
to affine transformations. If the equation f(x,y)=0 describes a shape, show that after transforming it with
transformation T, the new shape is described by all points that satisfy g(x,y) = f(T-1(x,y)) = 0. Then show the
details of this form when T is affine. Finally, try to describe the largest class of shapes which is preserved under
affine transformations.

Case Study 5. 4. Generalized 3D Shears.

Chap 5. Transformations

9/28/99

page 64

(Level of Effort:II) A shear can be more general than those discussed in Section 5.???. As suggested by Goldman
[goldman 91] the ingredients of a shear are:
A plane through the origin having unit normal vector m;
A unit vector v lying in the plane (thus perpendicular to m);
An angle .
Then as shown in Figure 5.70 point P is sheared to point Q by shifting it in direction v a certain amount. The
amount is proportional to both the distance at which P lies from the plane, and to tan . Goldman shows that this
shear has the matrix representation:
m

P
Q

P0
v

Figure 5.70. Defining a shear in 3D.

mx v x
M = I + tan( ) my v x

mz vx

m xv y
m yv y
mz v y

m x vz
m yv z

m zv z

(5.44)

where I is the 3 by 3 identity matrix, and an offset vector of 0. Some details of the derivation are given in the
exercises.
Example 5.8.2: Find the shear associated with the plane having unit normal vector m = (1,1,1)/ 3 = (0.577,
0.577, 0.577), unit vector v = (0, 0.707 -0.707), and angle = 30. Solution: Note that v lies in the plane as
required (why?). Applying Equation 9.9.5 we obtain:

0 0.408 0.408 1 0.235 0.235

M = I + 0.577 0 0.408 0.408 = 0 1.235 0.235

0 0.408 0.408 0 0.235 0.764

Derive the shear matrix. We want to express the point Q in Figure 5.73 in terms of P and the ingredients of the
shear. The (signed) distance of P from the plane is given by Pm (considering P as a position vector pinned to
the origin). a). Put the ingredients together to obtain:
Q = P + (Pm) tan() v

(5.45)

Note that points on the other side of the plane are sheared in the opposite direction, as we would expect.
Now we need to the second term as P times some matrix. Use Appendix 2 to show that Pm is PmT, and
therefore that Q = P ( I + tan() mT v), and that the shear matrix is ( I + tan() mT v). Show that the matrix mT v
has the form:

m xv x
mx
m v = m y (vx ,vy ,vz ) = my v x

mz
mz vx
T

Chap 5. Transformations

mx v y mx v z
m y v y my vz

mz vy m zvz

9/28/99

(5.46)

page 65

This is called the outer product (or tensor product) of m with v.

Practice Exercises.
5.8.9. Consistency with a simple shear. Express the matrix for the shear associated with direction v = (1,0,0)
and the plane having normal vector (0, 1, 0). Show that this is a form of the elementary shear matrix given in
Equation 5.45, for any angle .
5.8.10. Find the shear. Compute the shear matrix for the shear that uses a normal m = 0.577(1, -1, 1), direction
vector v = 0.707(1,1,0), and angle = 45. Sketch the vectors and plane involved, and sketch how a cube
centered at the origin with edges aligned with the coordinate axes is affected by the shear.
5.8.11. How Is a Three-Dimensional Shear Like a Two-Dimensional Translation in Homogeneous
Coordinates?
Consider the specific shear:

1
Q = P 0
t

0 0
1 0

s 1

(5.47)

which alters the 3D point P by shearing it along x with factor t owing to z and along y with factor s owing to z. If
P lies in the z = 1 plane with coordinates P = (px, py, 1), show that P is transformed into (px + t, py + s, 1).
Hence it is simply shifted in x by amount t and in y by amount s. Show that for any point in the z = 1 plane, this
shear is equivalent to a translation. Show that the matrix in Equation 5.48 is identical to the homogeneous
coordinate form for a pure translation in two dimensions. This correspondence helps reinforce how homogeneous
coordinates operate.

Case Study 5.5. Rotation about an Axis: The Constructive Approach.

(Level of Effort: II) You are asked to fill in the details in the following derivation of the rotation matrix Ru() of
Equation 9.9.15. For simplicity we assume u is a unit vector: |u| = 1. Denote the position vector based on P by
the name p, so p = P - O where O is the origin of the coordinate system. Now project p onto u to obtain vector h,
as shown in Figure 9.9.8a.
a). Show it has the form h = (p u) u. Define two perpendicular vectors a and b that lie in the plane of rotation: a
= p - h, and b = u a
b). Show that they are perpendicular, they have the same length, they both lie in the plane, and that b = u (p h) simplifies to u p. This effectively establishes a 2D coordinate system in the plane of rotation. Now look
onto the plane of rotation, as in Figure 9.9.8b. The rotation rotates a to a = cos a + sin b, so the rotated point
Q is given by:
Q = h + cos a + sin b, or using the expressions for a and b,
Q = p cos + (1 - cos ) (p u) u + sin (u p)

(5.48)

This is a rather general result, showing how the rotated point Q can be decomposed into portions along h and
along two orthogonal axes lying in the plane of rotation.
This form for Q hardly looks like the multiplication of P by some matrix, but it is because each of the three terms
is linear in p. Convert each of the terms to the proper form as follows:
c). Replace p with P to obtain immediately p (cos ) = P (cos ) I where I is the 3 by 3 identity matrix.
d). Use the result (Appendix 2) that a dot product p c can be rewritten as P times a matrix: P uT, to show that
(p u)u = P uTu where uTu is an outer product similar to Equation 9.9.7.
e). Use the fact developed in Appendix 2 that a cross product u p can also be written as P times a matrix to
show that u p = P Cross(u) where matrix Cross(c) is:

Chap 5. Transformations

9/28/99

page 66

0
Cross(u) = uz

uz
0
ux

uy
ux

(5.49)

f). Put these together to obtain the matrix

(5.50)17

Ru() = cos I + (1 - cos ) uTu + sin Cross(u))

M is therefore the sum of three weighted matrices. It surely is easier to build than the product of five matrices as
in the classic route.
g). Write this out to obtain Equation 5.33.

Case Study 5.6. Decomposing 3D Affine Transformations.

(Level of Effort:III) This Case Study looks at several broad families of affine tansformations.
What is in a 3D Affine Transformation?
Once again we ignore the translation portion of an affine transformation and focus on the linear transformation
part represented by 3 x 3 matrix M. What kind of transformations are imbedded in it? It is somewhat more
complicated. Goldman[GEMSIII, p.108] shows that every such M is the product of a scaling S, a rotation R , and
two shears H1 and H2.
M = S R H1 H2

(5.51)

Every 3D affine transformation, then, can be viewed as this sequence of elementary operations, followed by a
translation. In Case Study 5.??? we explore the mathematics behind this form, and see how an actual
decomposition might be carried out.
Useful Classes of Transformations.
Its useful to categorize affine transformations, according to what they do or dont do to certain properties of
an object when it is transformed. We know they always preserve parallelism of edges of an object, but which
transformations also preserve the length of each edge, and which ones preserve the angles between each pair of
edges?
1). Rigid Body Motions.
It is intuitively clear that translating an object, or rotating it, will not change its shape or size. In addition,
reflecting it about a plane has no effect on its shape or size. Since the shape is not affected by any one of these
transformations alone, it is also not affected by an arbitrary composition of them. We denote by
Trigid = {rotations, reflections, translations}
the collection of all affine transformations that consist of any sequence of rotations, reflections, and translations.
These are known classically as the rigid body motions, since an object is moved rigidly from one position and
orientation to another. Such transformations have orthogonal matrices in homogeneous coordinates. These are
matrices for which the inverse is the same as the transpose:

~
~
M 1 = M T
2). Angle-Preserving Transformations.
A uniform scaling (having equal scale factors Sx = Sy = Sz) expands or shrinks an object, but does so uniformly,
so there is no change in the objects shape. Thus the angle between any two edges is unaffected. We can denote
such a class as

17Goldman [gold90] reports the same form for M, and gives compact results for several other complex transformations.

Chap 5. Transformations

9/28/99

page 67

Tangle = {rotations, reflections, translations, uniform scalings}

This class is larger than the rigid body motions, because it includes uniform scaling. It is an important class
because, as we see in Chapter ???, lighting and shading calculations depend on the dot products between various
vectors. If a certain transformation does not alter angles then it does not alter dot products, and lighting
calculations can take place in either the transformed or the untransformed space.
(Estimate of time required: three hours.) Given a 3D affine transformation {M, d} we wish to see how it is
composed of a sequence of elementary transformations. Following Goldman [GEMS III, p. 108] we develop the
steps required to decompose the 3 by 3 matrix M into the product of a scaling S, a rotation R, and two shears H1
and H2:
M = S R H1 H2

(5.52)

You are asked to verify each step along the way, and to develop a routine that will produce the individual
matrices S,R , H1 and H2.
Suppose the matrix M has rows u, v, and w, each a 3D vector:

u
M = v

Goldmans approach is based on the classical Gram-Schmidt orthogonalization procedure, whereby the rows of
M are combined in such a way that they become mutually orthogonal and of unit length. The matrix composed of
these rows is therefore orthogonal (see Practice Exercise 9.9.14) and so represents a rotation (or a rotation with a
reflection). Goldman shows that the orthogonalization process is in fact two shears. The rest is detail.
The steps are as follows. Carefully follow each step, and do each of the tasks given in brackets.
1. Normalize u to u* = u/S1 , where S1 = |u|.
2. Subtract a piece of u* from v so that what is left is orthogonal to u*: Call b = v - du* where d = vu*. [Show
that bu* = 0.]
3. Normalize b: set v* = b/S2, where S2 = |b|.
4. Set up some intermediate values: m = wu* and n = wv*, e = m + n , and r = (mu* + nv*)/e.
5. Subtract a piece of r from w so that what is left is orthogonal to both u* and v*: Call c = w - er. [Show that
cu* = cv* = 0.]
5. Normalize c: set w* = c/S3 where S3 = |c|.
7. The matrix
2

u *
R = v *
w *

is therefore orthogonal, and so represents a rotation. (Compute its determinant: if it is -1 then simply replace w*
with -w*.)
8. Define the shear matrix H1 = I +

d
(v * u*) (see Practice exercise 9.9.1), where
S2

(v * u) = (v) T u * is the outer product of v* and u*.

e
(w * r) , where (w * r) = (w*)T r .
9. Define the shear matrix H2 = I +
S3
10. [Show that u*H1 = u*, v*H1 = v* + du*/S2 = v/S2, and w* H1 = w*. First show the property of the outer
product that for any vectors a, b, and c: a(b c) = (a b)c . Then use this property, along with the
orthogonality of u*, v* and w*, to show the three relations.]
11. [Show that u*H2 = u*, v*H2 = v*, w*H2 = w* + er/ S3 = w/S3.

Chap 5. Transformations

9/28/99

page 68

12. [Put these together to show that M = SRH1H2.] S is defined to be

S1
S = 0
0

0
S2
0

0
0

Note that the decomposition is not unique, since the vectors u, v, and w could be orthogonalized in a different
order. For instance, we could first form w* as w/|w|, then subtract a piece of v from w* to make a vector
orthogonal to w*, then subtract a vector from u to make a vector orthogonal to the other two.
Write the routine:
void decompose(DBL m[3][3],DBL S[3][3],DBL R[3][3],DBL H1[3][3],DBL
H2[3][3])
where DBL is defined as double, that takes matrix m and computes the matrices S, R, H1 and H2 as described
above. Test your routine on several matrices.
Other ways to decompose a 3D transformation have been found as well. See for instance [thomas, GEMS II, p.
320] and [Shoemake, GEMS IV p. 207].

Case Study 5.7. Drawing 3D scenes described by SDL.

(Level of Effort:II) Develop a complete application that uses the Scene, Shape, Affine4, etc. classes and
supports reading in and drawing the scene described in an SDL file. To assist you, use any classes provided on
the books web site. Flesh out any drawOpenGL() methods that are needed. Develop a scene file that contains
descriptions of the jack, a wooden chair, and several walls.

5.9. For Further Reading

There are some excellent books on transformations that supplement the treatment here. George Martins
TRANSFORMATION GEOMETRY [martin82] is superb, as is Yagloms GEOMETRIC
TRANSFORMATIONS [yaglom62]. Hogar also provides a fine chapter on vectors and transformations
[hoggar92]. Several of Blinns engaging articles in JIM BLINNS CORNER - A TRIP DOWN THE GRAPHICS
PIPELINE [blinn96] provide excellent discussions of homogeneous coordinates and transformations in computer
graphics.

Chap 5. Transformations

9/28/99

page 69

(for ECE660, Fall, 1999 F.S.Hill,Jr.)

CHAP 6. Modeling Shapes with Polygonal Meshes

Try to learn something about everything and everything about something.
T.H. Huxley
Goals of the Chapter
To develop tools for working with objects in 3D space
To represent solid objects using polygonal meshes
To draw simple wireframe views of mesh objects
Preview
Section 6.1 gives an overview of the 3D modeling process. Section 6.2 describes polygonal meshes that allow
you to capture the shape of complex 3D objects in simple data structures. The properties of such meshes are
reviewed, and algorithms are presented for viewing them.
Section 6.3 describes families of interesting 3D shapes, such as the Platonic solids, the Buckyball, geodesic
domes, and prisms. Section 6.4 examines the family of extruded or swept shapes, and shows how to create
3D letters for flying logos, tubes and snakes that undulate through space, and surfaces of revolution.
Section 6.5 discusses building meshes to model solids that have smoothly curved surfaces. The meshes approximate some smooth underlying surface. A key ingredient is the computation of the normal vector to the
underlying surface at any point. Several interesting families of smooth solids are discussed, including quadric
and superquadric surfaces, ruled surfaces and Coons patches, explicit functions of two variables, and surfaces
of revolution.
The Case Studies at the end of the chapter explore many ideas further, and request that you develop applications to test things out. Some cases studies are theoretical in nature, such as deriving the Newell formula for
computing a normal vector, and examining the algebraic form for quadric surfaces. Others are more practical,
such as reading mesh data from a file and creating beautiful 3D shapes such as tubes and arches.

6.1 Introduction.
In this chapter we examine ways to describe 3D objects using polygonal meshes. Polygonal meshes are simply
collections of polygons, or faces, which together form the skin of the object. They have become a standard
way of representing a broad class of solid shapes in graphics. We have seen several examples, such as the
cube, icosahedron, as well as approximations of smooth shapes like the sphere, cylinder, and cone (see Figure
5.65). In this chapter we shall see many more examples. Their prominence in graphics stems from the simplicity of using polygons. Polygons are easy to represent (by a sequence of vertices) and transform, have simple
properties (a single normal vector, a well-defined inside and outside, etc.), and are easy to draw (using a polygon-fill routine, or by mapping texture onto the polygon).
Many rendering systems, including OpenGL, are based on drawing objects by drawing a sequence of polygons.
Each polygonal face is sent through the graphics pipeline (recall Figure 5.56), where its vertices undergo various transformations, until finally the portion of the face that survives clipping is colored in, or shaded, and
shown on the display device.
We want to see how to design complicated 3D shapes by defining an appropriate set of faces. Some objects
can be perfectly represented by a polygonal mesh, whereas others can only be approximated. The barn of Figure 6.1a, for example, naturally has flat faces, and in a rendering the edges between faces should be visible.
But the cylinder in Figure 6.1b is supposed to have a smoothly rounded wall. This roundness cannot be
achieved using only polygons: the individual flat faces are quite visible, as are the edges between them. There
are, however, rendering techniques that make a mesh like this appear to be smooth, as in Figure 6.1c. We examine the details of so-called Gouraud shading in Chapter 8.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 1

Figure 6.1. Various shapes modeled by meshes.

We begin by describing polygonal meshes in some generality, and seeing how to define and manipulate them
in a program (with or without OpenGL). We apply these ideas to modeling polyhedra, which inherently have
flat faces, and study a number of interesting families of polyhedra. We then tackle the problem of using a
mesh to approximate a smoothly curved shape, and develop the necessary tools to create and manipulate such
models.

6.2. Introduction to Solid Modeling with Polygonal Meshes.

I never forget a face, but in your case I'll make an exception.
Grouch Marx
We shall use meshes to model both solid shapes and thin skins. The object is considered to be solid if the
polygonal faces fit together to enclose space. In other cases the faces fit together without enclosing space, and
so they represent an infinitesimally thick surface. In both cases we call the collection of polygons a polygonal
mesh (or simply mesh).
A polygonal mesh is given by a list of polygons, along with information about the direction in which each
polygon is facing. This directional information is often simply the normal vector to the plane of the face,
and it is used in the shading process to determine how much light from a light source is scattered off the face.
Figure 6.2 shows the normal vectors for various faces of the barn. As we examine in detail in Chapter 11, one
component of the brightness of a face is assumed to be proportional to the cosine of the angle (shown as in
the figure for the side wall of the barn) between the normal vector to a face and the vector to the light source.
Thus the orientation of a surface with respect to light sources plays an important part in the final drawing.

Figure 6.2. The normal direction to a face determines its brightness.

Vertex normals versus face normals.
It turns out to be highly advantageous to associate a normal vector to each vertex of a face, rather than one
vector to an entire face. As we shall see, this facilitates the clipping process as well as the shading process for
smoothly curved shapes. For flat surfaces such as the wall of a barn, this means that each of the vertices V1, V2,
V3, and V4 that define the side wall of the barn will be associated with the same normal n1, the normal vector to
the side wall itself (see Figure 6.3a). But vertices of the front wall, such as V5, will use normal n2. (Note that
vertices V1 and V5 are located at the same point in space, but use different normals.)

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 2

Figure 6.3. Associating a normal with each vertex of each face.

For a smoothly curved surface such as the cylinder in Figure 6.3b, a different approach is used in order to permit shading that makes the surface appear smooth. Both vertex V1 of face F1 and vertex V2 on face F2 use the
same normal n, which is the vector perpendicular to the underlying smooth surface. We see how to compute
this vector conveniently in Section 6.2.2.
6.2.1. Defining a polygonal mesh.
A polygonal mesh is a collection of polygons along with a normal vector associated with each vertex of each
polygon. We begin with an example.
Example 6.2.1. The basic barn. Figure 6.4 shows a simple shape we call the basic barn. It has seven polygonal faces and a total of 10 vertices (each of which is shared by three faces). For convenience it has a
square floor one unit on a side. (The barn would be scaled and oriented appropriately before being placed in a
scene.) Because the barn is assumed to have flat walls there are only seven distinct normal vectors involved,
the normal to each face as shown.

Figure 6.4. Introducing the basic barn.

There are various ways to store mesh information in a file or program. For the barn you could use a list of
seven polygons, and list for each one where its vertices are located and where the normal for each of its vertices is pointing (a total of 30 vertices and 30 normals). But this would be quite redundant and bulky, since there
are only 10 distinct vertices and seven distinct normals.
A more efficient approach uses three separate lists, a vertex list, normal list, and face list. The vertex list reports the locations of the distinct vertices in the mesh. The list of normals reports the directions of the distinct
normal vectors that occur in the model. The face list simply indexes into these lists. As we see next, the barn is
thereby captured with 10 vertices, seven normals, and a list of seven simple face descriptors.
The three lists work together: The vertex list contains locational or geometric information, the normal list
contains orientation information, and the face list contains connectivity or topological information.
The vertex list for the barn is shown in Figure 6.5. The list of the seven distinct normals is shown in Figure
6.6. The vertices have indices 0 through 9 and the normals have indices 0 through 6. The vectors shown have
already been normalized, since most shading algorithms require unit vectors. (Recall that a cosine can be
found as the dot product between two unit vectors).
vertex
x
y
z
0
0
0
0
1
0
0
1
1
1
0
2
0.5 1.5
0
3
0
1
0
4
0
0
1
5
1
0
1
6
1
1
1
7
0.5 1.5
1
8
0
1
1
9
Figure 6.5. Vertex list for the basic barn.
normal
0

nx
-1

ny
0

nz
0

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 3

-0.477 0.8944
0
1
0.447
0.8944
0
2
1
0
0
3
0
-1
0
4
0
0
1
5
0
0
-1
6
Figure 6.6. The list of distinct normal vectors involved.
face
vertices
0,5,9,4
0 (left)
3,4,9,8
1 (roof left)
2,3,8,7
2 (roof right)
1,2,7,6
3 (right)
0,1,6,5
4 (bottom)
5,6,7,8,9
5 (front)
0,4,3,2,1
6 (back)
Figure 6.7. Face list for the basic barn.

associated normal
0,0,0,0
1,1,1,1
2,2,2,2
3,3,3,3
4,4,4,4
5,5,5,5,5
6,6,6,6,6

Figure 6.7 shows the barns face list: each face has a list of vertices and the normal vector associated with each
vertex. To save space, just the indices of the proper vertices and normals are used. (Since each surface is flat
all of the vertices in a face are associated with the same normal.) The list of vertices for each face begins with
any vertex in the face, and then proceeds around the face vertex by vertex until a complete circuit has been
made. There are two ways to traverse a polygon: clockwise and counterclockwise. For instance, face #5 above
could be listed as (5,6,7,8,9) or (9,8,7,6,5). Either direction could be used, but we follow a convention that
proves handy in practice:
Traverse the polygon counterclockwise as seen from outside the object.
Using this order, if you traverse around the face by walking on the outside surface from vertex to vertex, the
interior of the face is on your left. We later design algorithms that exploit this ordering. Because of it the algorithms are able to distinguish with ease the front from the back of a face.
The barn is an example of a data intensive model, where the position of each vertex is entered (maybe by
hand) by the designer. In contrast, we see later some models that are generated algorithmically. Here it isnt
too hard to come up with the vertices for the basic barn: the designer chooses a simple unit square for the
floor, decides to put one corner of the barn at the origin, and chooses a roof height of 1.5 units. By suitable
scaling these dimension can be altered later (although the relative height of the wall to the barns peak, 1: 1.5,
is forever fixed).
6.2.2. Finding the normal vectors.
It may be possible to set vertex positions by hand, but it is not so easy to calculate the normal vectors. In general each face will have three or more vertices, and a designer would find it challenging to jot down the normal
vector. Its best to let the computer do it during the creation of the mesh model.
If the face is considered flat, as in the case of the barn, we need only find the normal vector to the face itself,
and associate it with each of the faces vertices. One direct way uses the vector cross product to find the normal, as in Figure 4.16. Take any three adjacent points on the face, V1, V2, and V3, and compute the normal as
their cross product m = (V1 - V2)X(V3 - V2). This vector can now be normalized to unit length.
There are two problems with this simple approach. a). If the two vectors V1 - V2 and V3 - V2 are nearly parallel
the cross product will be very small (why?), and numerical inaccuracies may result. b). As we see later, it may
turn out that the polygon is not perfectly planar, i.e. that all of the vertices do not lie in the same plane. Thus

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 4

the surface represented by the vertices cannot be truly flat. We need to form some average value for the
normal to the polygon, one that takes into consideration all of the vertices.
A robust method that solves both of these problems was devised by Martin Newell [newm79]. It computes the
components mx, my, mz of the normal m according to the formula:
N 1

m x = (yi ynext( i ) )(z i + znext (i ) )

i= 0
N 1

m y = (z i znext ( i) )(xi + x next( i ) )

(6.1)

i=0
N 1

m z = (xi xnext (i ) )(yi + ynext( i ) )

i =0

where N is the number of vertices in the face, (xi, yi, zi) is the position of the i-th vertex, and where next(j) =
(j+1) mod N is the index of the next vertex around the face after vertex j, in order to take care of the wraparound from the (N-1)-st to the 0-th vertex. The computation requires only one multiplication per edge for
each of the components of the normal, and no testing for collinearity is needed. This result is developed in
Case Study 6.2, and C++ code is presented for it.
The vector m computed by the Newell method could point toward the inside or toward the outside of the polygon. We also show in the Case study that if the vertices of the polygon are traversed (as i increases) in a CCW
direction as seen from outside the polygon, then m points toward the outside of the face.
Example 6.2.2: Consider the polygon with vertices P0 = (6, 1, 4), P1 = (7, 0, 9), and P2 = (1, 1, 2). Find the
normal to this polygon using the Newell method. Solution: Direct use of the cross product gives ((7, 0, 9) - (6,
1, 4)) ((1, 1, 2) - (6, 1, 4)) = (2, -23, -5). Application of the Newell method yields the same result: (2, -23,
-5).
Practice Exercises.
6.2.1. Using the Newell Method.
For the three vertices (6, 1, 4), (2, 0, 5), and (7, 0, 9), compare the normal found using the Newell method
with that found using the usual cross product. Then use the Newell method to find (nx,ny,nz) for the polygon
having the vertices (1, 1, 2), (2, 0, 5), (5, 1, 4), (6, 0, 7). Is the polygon planar? If so, find its true normal using
the cross product, and compare it with the result of the Newell method.
6.2.2. What about a non-planar polygon? Consider the quadrilateral shown in Figure 6.8 that has the vertices
(0, 0, 0), (1, 0, 0), (0, 0,1), and (1, a, 1). When a is nonzero this is a non-planar polygon. Find the normal to
it using the Newell method, and discuss how good an estimate it is for different values of a.
Figure 6.8. A non-planar polygon.
6.2.3. Represent the generic cube. Make vertex, normal, and face lists for the generic cube, which is
centered at the origin, has its edges aligned with the coordinate axes, and has edges of length two. Thus its
eight vertices lie at the eight possible combinations of + and - in (1, 1, 1).
6.2.4. Faces with holes. Figure 6.9 shows how a face containing a hole can be captured in a face list. A pair of
imaginary edges are added that bridge the gap between the circumference of the face and the hole, as suggested in the figure.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 5

2
7

6
9
4

8
3

Figure 6.9. A face containing a hole.

The face is traversed so that (when walking along the outside surface) the interior of the face lies to the left.
Thus a hole is traversed in the CW direction. Assuming we are looking at the face in the figure from its outside, the list of vertices would be: 5 4 3 8 9 6 7 8 3 2 1. Sketch this face with an additional hole in it, and give
the proper list of vertices for the face. What normals would be associated with each vertex?
6.2.3. Properties of Meshes.
Given a mesh consisting of a vertex, normal, and face lists, we might wonder what kind of an object it represents. Some properties of interest are:
Solidity: As mentioned above a mesh represents a solid object if its faces together enclose a positive and
finite amount of space.
Connectedness: A mesh is connected if every face shares at least one edge with some other face. (If a mesh
is not connected it is usually considered to represent more than one object.)
Simplicity: A mesh is simple if the object it represents is solid and has no holes through it; it can be deformed into a sphere without tearing. (Note that the term simple is being used here in quite a different sense
from for a simple polygon.).
Planarity: A mesh is planar if every face is a planar polygon: The vertices of each face then lie in a single
plane. Some graphics algorithms work much more efficiently if a face is planar. Triangles are inherently planar, and some modeling software takes advantage of this by using only triangles. Quadrilaterals, on the other
hand, may or may not be planar. The quadrilateral in Figure 6.8, for instance, is planar if and only if a = 0.
Convexity: The mesh represents a convex object if the line connecting any two points within the object lies
wholly inside the object. convexity was first discussed in section 2.3.6 in connection with polygons. Figure
6.10 shows some convex and some non-convex objects. For each non-convex object an example line is shown
whose endpoints lie in the object but which is not itself contained within the object.

Figure 6.10. Examples of convex and nonconvex 3D objects.

The basic barn happens to possess all of these properties (check this). For a given mesh some of these properties are easy to determine in a program, in the sense that a simple algorithm exists that does the job. (We discuss some below.) Other properties, such as solidity, are quite difficult to establish.
The polygonal meshes we choose to use to model objects in graphics may have some or all of these properties:
the choice depends on how one is going to use the mesh. If the mesh is supposed to represent a physical object
made of some material, perhaps to determine its mass or center of gravity, we may insist that it be at least

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 6

connected and solid. If we just want to draw the object, however, much greater freedom is available, since
many objects can still be drawn, even if they are non-physical.
Figure 6.11 shows some examples of objects we might wish to represent by meshes. PYRAMID is made up of
triangular faces which are necessarily planar. It is not only convex, in fact it has all of the properties above.
DONUT
PYRAMID

IMPOSSIBLE

BARN

Figure 6.11. Examples of solids to be described by meshes.

DONUT is connected and solid, but it is neither simple (there is a hole through it) nor convex. Two of its faces
themselves have holes. Whether or not its faces are planar polygons cannot be determined from a figure alone.
Later we give an algorithm for determining planarity from the face and vertex lists.
IMPOSSIBLE cannot exist in space. Can it really be represented by a mesh?
BARN seems to be made up of two parts, but it could be contained in a single mesh. Whether it is connected
would then depend on how the faces were defined at the juncture between the silo and the main building.
BARN also illustrates a situation often encountered in graphics. Some faces have been added to the mesh that
represent windows and doors, to provide texture to the object. For instance, the side of the barn is a rectangle with no holes in it, but two squares have been added to the mesh to represent windows. These squares are
made to lie in the same plane as the side of the barn. This then is not a connected mesh, but it can still be displayed on a graphics device.
6.2.4. Mesh models for non-solid objects.
Figure 6.12 shows examples of other objects that can be characterized by polygonal meshes. These are surfaces, and are best thought of as infinitesimally thick shells.
a).
b).
c).
BOX
STRUCT

FACE

(use b), c) from old Fig. 15.3)

Figure 6.12. Some surfaces describable by meshes. (Part c is courtesy of the University of Utah.)
BOX is an open box whose lid has been raised. In a graphics context we might want to color the outside of
BOX's six faces blue and their insides green. (What is obtained if we remove one face from PYRAMID
above?)

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 7

Two complex surfaces, STRUCT and FACE, are also shown. For these examples the polygonal faces are being
used to approximate a smooth underlying surface. In some situations the mesh may be all that is available for
the object, perhaps from digitizing points on a person's face. If each face of the mesh is drawn as a shaded
polygon, the picture will look artificial, as seen in FACE. Later we shall examine tools that attempt to draw
the smooth underlying surface based only on the mesh model.
Many geometric modeling software packages construct a model from some object - a solid or a surface - that
tries to capture the true shape of the object in a polygonal mesh. The problem of composing the lists can be
difficult. As an example, consider creating an algorithm that generates vertex and face lists to approximate the
shape of an engine block, a prosthetic limb, or a building. This area in fact is a subject of much ongoing research [Mant88], [Mort85]. By using a sufficient number of faces, a mesh can approximate the underlying
surface to any degree of accuracy desired. This property of completeness makes polygon meshes a versatile
tool for modeling.
6.2.5. Working with Meshes in a Program.
We want an efficient way to capture a mesh in a program that makes it easy to create and draw the object.
Since mesh data is frequently stored in a file, we also need simple ways to read and write mesh files.
It is natural to define a class Mesh, and to imbue it with the desired functionality. Figure 6.13 shows the declaration of the class Mesh, along with those of two simple helper classes, VertexID and Face1. A Mesh object has a vertex list, a normal list, and a face list, represented simply by arrays pt, norm, and face, respectively. These arrays are allocated dynamically at runtime, when it is known how large they must be. Their
lengths are stored in numVerts, numNormals, and numFaces, respectively. Additional data fields can be
added later that describe various physical properties of the object such as weight and type of material.
//################# VertexID ###################
class VertexID{
public:
int vertIndex; // index of this vert in the vertex list
int normIndex; // index of this vertexs normal
};
//#################### Face ##################
class Face{
public:
int nVerts; // number of vertices in this face
VertexID * vert; // the list of vertex and normal indices
Face(){nVerts = 0; vert = NULL;} // constructor
~Face(){delete[] vert; nVerts = 0;} // destructor
};
//###################### Mesh #######################
class Mesh{
private:
int numVerts;
// number of vertices in the mesh
Point3* pt;
// array of 3D vertices
int numNormals;
// number of normal vectors for the mesh
Vector3 *norm;
// array of normals
int numFaces;
// number of faces in the mesh
Face* face;
// array of face data
// ... others to be added later
public:
Mesh();
// constructor
~Mesh();
// destructor
int readFile(char * fileName); // to read in a filed mesh
.. others ..
};
Figure 6.13. Proposed data type for a mesh.
1

Definitions of the basic classes Point3 and Vertex3 have been given previously, and also appear in Appendix 3.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 8

The Face data type is basically a list of vertices and the normal vector associated with each vertex in the face.
It is organized here as an array of index pairs: the v-th vertex in the f-th face has position pt[face[f].
vert[v].vertIndex] and normal vector norm[face[f].vert[v].normIndex]. This appears
cumbersome at first exposure, but the indexing scheme is quite orderly and easy to manage, and it has the advantage of efficiency, allowing rapid random access indexing into the pt[] array.
Example 6.2.3. Data for the tetrahedron. Figure 6.14 shows the specific data structure for the tetrahedron
shown, which has vertices at (0,0,0), (1,0,0), (0,1,0), and (0,0,1). Check the values reported in each field. (We
discuss how to find the normal vectors later.)

Figure 6.14. Data for the tetrahedron.

The first task is to develop a method for drawing such a mesh object. Its a matter of drawing each of its faces,
of course. An OpenGL-based implementation of the Mesh::draw() method must traverse the array of faces in
the mesh object, and for each face send the list of vertices and their normals down the graphics pipeline. In
OpenGL you specify that subsequent vertices are associated with normal vector m by executing glNormal3f(m.x, m.y, m.z)2. So the basic flow of Mesh::draw() is:
for(each face, f, in the mesh)
{
glBegin(GL_POLYGON);
for(each vertex, v, in face, f)
{
glNormal3f(normal at vertex v);
glVertex3f(position of vertex v);
}
glEnd();
}
The implementation is shown in Figure 6.15.
void Mesh:: draw() // use OpenGL to draw this mesh
{
for(int f = 0; f < numFaces; f++) // draw each face
{
glBegin(GL_POLYGON);
for(int v = 0; v < face[f].nVerts; v++) // for each one..
{
int in = face[f].vert[v].normIndex ; // index of this normal
int iv = face[f].vert[v].vertIndex ; // index of this vertex
glNormal3f(norm[in].x, norm[in].y, norm[in].z);
glVertex3f(pt[iv].x, pt[iv].y, pt[iv].z);
}
glEnd();
}
}
Figure 6.15. Method to draw a mesh using OpenGL.
We also need methods to create a particular mesh, and to read a predefined mesh into memory from a file. We
consider reading from and writing to files in Case Study 6.1. We next examine a number of interesting families of shapes that can be stored in a mesh, and see how to create them.
2

For proper shading these vectors must be normalized. Otherwise place glEnable(GL_NORMALIZE) in the
init() function. This requests that OpenGL automatically normalize all normal vectors.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 9

Creating and drawing a Mesh object using SDL.

It is convenient in a graphics application to read in mesh descriptions using the SDL language introduced in
Chapter 5. To make this possible simply derive the Mesh class from Shape, and add the method drawOpenGL(). Thus Figure 6.13 becomes:
class Mesh : public Shape {
//same as in Figure 6.13
virtual void drawOpenGL
{
tellMaterialsGL(); glPushMatrix(); // load properties into the object
glMultMatrixf(transf.m); draw(); // draw the mesh
glPopMatrix();
}
}; //end of Mesh class
The Scene class that reads SDL files is already set to accept the keyword mesh, followed by the name of the
file that contains the mesh description. Thus to create and draw a pawn with a certain translation and scaling,
use:
push translate 3 5 4 scale 3 3 3 mesh pawn.3vn pop

6.3. Polyhedra.
It is frequently convenient to restrict the data in a mesh so that it represents a polyhedron. A very large number of solid objects of interest are indeed polyhedra, and algorithms for processing a mesh can be greatly simplified if they need only process meshes that represent a polyhedron.
Slightly different definitions for a polyhedron are used in different contexts, [Coxeter69, Courant & Robbins??, F&VD90,] but we use the following:
Definition
A polyhedron is a connected mesh of simple planar polygons that encloses a finite amount of space.
So by definition a polyhedron represents a single solid object. This requires that:
every edge is shared by exactly two faces;
at least three edges meet at each vertex:
faces do not interpenetrate: Two faces either do not touch at all, or they touch
only along their common edge.
In Figure 6.11 PYRAMID is clearly a polyhedron. DONUT evidently encloses space, so it is a polyhedron if
its faces are in fact planar. It is not a simple polyhedron, since there is a hole through it. In addition, two of its
faces themselves have holes. Is IMPOSSIBLE a polyhedron? Why? If the texture faces are omitted BARN
might be modeled as two polyhedra, one for the main part and one for the silo.
Eulers formula:
Eulers formula (which is very easy to prove; see for instance [courant and robbins, 61, p. 236]) provides a
fundamental relationship between the number of faces, edges, and vertices (F, E, and V respectively) of a simple polyhedron:
V+F-E= 2

(6.1)

For example, a cube has V = 8, F = 6, and E = 12.

A generalization of this formula is available if the polyhedron is not simple [F&VD p. 544]:
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 10

V + F - E = 2 + H - 2G

(6.2)

where H is the total number of holes occurring in faces, and G is the number of holes through the polyhedron.
Figure 6.16a shows a donut constructed with a hole in the shape of another parallelepiped. The two ends are
beveled off as shown. For this object V = 16, F = 16, E = 32, H = 0, and G = 1. Figure 6.16b shows a polyhedron having a hole A penetrating part way into it, and hole B passing through it.
a).
b).

Figure 6.16. Polyhedron with holes.

Here V = 24, F = 15, E = 36, H = 3, and G = 1. These values satisfy the formula above in both cases.
The structure of a polyhedron is often nicely revealed by a Schlegel diagram. This is based on a view of the
polyhedron from a point just outside the center of one of its faces, as suggested in Figure 6.17a. Viewing a
cube in this way produces the Schlegel diagram shown in Figure 6.17b. The front face appears as a large polygon surrounding the rest of the faces.
a).
b).

Figure 6.17. The Schlegel diagrams for a cube.

Figure 6.18 shows further examples. Part a) shows the Schlegel diagram of PYRAMID shown in Figure 6.11,
and parts b) and c) show two quite different Schlegel diagrams for the basic barn. (Which faces are closest to
the eye?).
a).
b).
c).

Figure 6.18. Schlegel diagrams for PYRAMID and the basic barn.
6.3.1. Prisms and Antiprisms.
A prism is a particular type of polyhedron that embodies certain symmetries, and therefore is quite simple to
describe. As shown in Figure 6.19 a prism is defined by sweeping (or extruding) a polygon along a straight
line, turning a 2D polygon into a 3D polyhedron. In Figure 6.19a polygon P is swept along vector d to form
the polyhedron shown in part b. When d is perpendicular to the plane of P the prism is a right prism. Figure
6.19c shows some block letters of the alphabet swept to form prisms.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 11

a).

b).

c).

Figure 6.19. Forming a prism.

A regular prism has a regular polygon for its base, and squares for its side faces. A hexagonal version is
shown in Figure 6.20a. A variation is the antiprism, shown in Figure 6.20b. This is not an extruded object.
Instead the top n-gon is rotated through 180 / n degrees and connected to the bottom n-gon to form faces that
are equilateral triangles. (How many triangular faces are there?) The regular prism and antiprism are examples
of semi-regular polyhedra, which we examine further later.
n-gon
n-gon
a).
b).

equilateral
triangle

square
Figure 6.20. A regular prism and antiprism.

Practice Exercises
6.3.1. Build the lists for a prism.
Give the vertex, normal, and face lists for the prism shown in Figure 6.21. Assume the base of the prism (face
#4) lies in the xy-plane, and vertex 2 lies on the z-axis at z = 4. Further assume vertex 5 lies 3 units along the xaxis, and that the base is an equilateral triangle.
2

#1 (top)
1

#2 (back)

#3 (front)

#5 (right)
6
5

#4 (bottom)

Figure 6.21. An example polyhedral object.

6.3.2. The unit cube. Build vertex, normal, and face lists for the unit cube: the cube centered at the origin
with its edges aligned with the coordinate axes. Each edge has length 2.
6.3.3. An Antiprism. Build vertex, normal, and face lists for an anti-prism having a square as its top polygon.
6.3.4. Is This Mesh Connected? Is the mesh defined by the list of faces: (4, 1, 3), (4, 7, 2, 1), (2, 7, 5), (3, 4,
8, 7, 9) connected? Try to sketch the object, choosing arbitrary positions for the nine vertices. What algorithm
might be used to test a face list for connectedness?
6.3.5. Build Meshes. For the DONUT, IMPOSSIBLE, and BARN objects in Figure 6.11, assign numbers to
each vertex and then write a face list.
6.3.6. Schlegel diagrams. Draw Schlegel diagrams for the prisms of Figure 6.20 and Figure 6.21.
6.3.2. The Platonic Solids.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 12

If all of the faces of a polyhedron are identical and each is a regular polygon, the object is a regular polyhedron. These symmetry constraints are so severe that only five such objects can exist, the Platonic solids3
shown in Figure 6.22 [Coxe61]. The Platonic solids exhibit a sublime symmetry and a fascinating array of
properties. They make interesting objects of study in computer graphics, and often appear in solid modeling
CAD applications.
a). Tetrahedron

d). Icosahedron

b). Hexahedron (cube)

c). Octahedron

e). Dodecahedron

Figure 6.22. The five Platonic solids.

Three of the Platonic solids have equilateral triangles as faces, one has squares, and the dodecahedron has
pentagons. The cube is a regular prism, and the octahedron is an antiprism (why?). The values of V, F, and E
for each solid are shown in Figure 6.23. Also shown is the Schlfli4 symbol (p,q) for each solid. It states that
each face is a p-gon and that q of them meet at each vertex.
Schlafli
solid
V
F
E
Tetrahedron
4
4
6
(3, 3)
Hexahedron
8
6
12
(4, 3)
Octahedron
6
8
12
(3, 4)
Icosahedron
12
20
30
(3, 5)
Dodecahedron
20
12
30
(5, 3)
Figure 6.23. Descriptors for the Platonic solids.
It is straightforward to build mesh lists for the cube and octahedron (see the exercises). We give example vertex and face lists for the tetrahedron and icosahedron below, and we discuss how to compute normal lists. We
also show how to derive the lists for the dodecahedron from those of the icosahedron, making use of their duality which we define next.
Dual polyhedra.
Each of the Platonic solids P has a dual polyhedron D. The vertices of D are the centers of the faces of P, so
edges of D connect the midpoints of adjacent faces of P. Figure 6.24 shows the dual of each Platonic solid
inscribed within it.
old A8.2 showing duals.
Figure 6.24. Dual Platonic solids.
The dual of a tetrahedron is also a tetrahedron.
The cube and octahedron are duals.
3Named

in honor of Plato (427-347 bc) who commemorated them in his Timaeus. But they were known before
this: a toy dodecahedron was found near Padua in Etruscan ruins dating from 500 BC.
4 Named for L. Schlfli, (1814-1895), a Swiss Mathematician.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 13

The icosahedron and dodecahedron are duals.

Duals have the same number, E, of edges, and V for one is F for the other. In addition, if (p, q) is the Schlafli
symbol for one, then it is (q, p) for the other.
If we know the vertex list of one Platonic solid P, we can immediately build the vertex list of its dual D, since
vertex k of D lies at the center of face k of P. Building D in this way actually builds a version that inscribes P.
To keep track of vertex and face numbering we use a model, which is fashioned by slicing along certain edges
of each solid and unfolding it to lie flat, so that all the faces are seen from the outside. Models for three of
the Platonic solids are shown in Figure 6.25.
a). Tetrahedron

b). Cube
3

2
4

6
6

1
2

3
3

2
2

7
1

4
3

5
2

a). Octahedron
4

2
4

5
5

1
3

3
7

Figure 6.25. Models for the tetrahedron, cube, and octahedron.

Consider the dual pair of the cube and octahedron. Face 4 of the cube is seen to be surrounded by vertices 1, 5,
6, and 2. By duality vertex 4 of the octahedron is surrounded by faces 1, 5, 6, and 2. Note for the tetrahedron
that since it is self-dual, the list of vertices surrounding the k-th face is identical to the list of faces surrounding
the k-th vertex.
The position of vertex 4 of the octahedron is the center of face 4 of the cube. Recall from Chapter 5 that the
center of a face is just the average of the vertices belonging to that face. So if we know the vertices V1, V5, V6,
and V2 for face 4 of the cube we have immediately that:
1
V4 = 4 (V1 + V5 + V6 + V2 )

(6.3)

Practice Exercises
6.3.7. The octahedron. Consider the octahedron that is the dual of the cube described in Exercise 9.4.3. Build
vertex and face lists for this octahedron.
6.3.8. Check duality. Beginning with the face and vertex lists of the octahedron in the previous exercise, find
its dual and check that this dual is (a scaled version of) the cube.
Normal Vectors for the Platonic Solids.
If we wish to build meshes for the Platonic solids we must compute the normal vector to each face. This can
be done in the usual way using Newells method, but the high degree of symmetry of a Platonic solid offers a
much simpler approach. Assuming the solid is centered at the origin, the normal vector to each face is the
vector from the origin to the center of the face, which is formed as the average of the vertices. Figure 6.26
shows this for the octahedron. he normal to the face shown is simply:
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 14

Figure 6.26. Using symmetry to find the normal to a face.

m = (V1 + V2 + V3) / 3

(6.4)

(Note also: This vector is also the same as that from the origin to the appropriate vertex on the dual Platonic
solid.)
The Tetrahedron
The vertex list of a tetrahedron depends of course on how the tetrahedron is positioned, oriented, and sized. It
is interesting that a tetrahedron can be inscribed in a cube (such that its four vertices lie in corners of the cube,
and its four edges lie in faces of the cube). Consider the unit cube having vertices (1,1,1), and choose the
tetrahedron that has one vertex at (1,1,1). Then it has vertex and face lists given in Figure 6.27 [blinn87].
face list
vertex list
vertex
x
y
z
face number
vertices
0
1
1
1
0
1,2,3
1
1
-1
-1
1
0,3,2
2
-1
-1
1
2
0,1,3
3
-1
1
-1
3
0,2,1
Figure 6.27. Vertex List and Face List for a Tetrahedron.
The Icosahedron
The vertex list for the icosahedron presents more of a challenge, but we can exploit a remarkable fact to make
it simple. Figure 6.28 shows that three mutually perpendicular golden rectangles inscribe the icosahedron, and
so a vertex list may be read directly from this picture. We choose to align each golden rectangle with a coordinate axis. For convenience, we size the rectangles so their longer edge extends from -1 to 1 along its axis. The
shorter edge then extends from - to , where = ( 5 1) / 2 = 0.618... is the reciprocal of the golden
ratio .
old Fig A8.4 golden rectangles in icosahedron.
Figure 6.28. Golden rectangles defining the icosahedron.
From this it is just a matter of listing vertex positions, as shown in Figure 6.29.
x
y
z
vertex
0
0
1

1
0
1
-
2
1
0

3
1
- 0
4
0
-1 -
5
0
-1
6
0
1

7
0
1
-
8
0
-1

9
0
-1
-
10
-1
0

11
-1
- 0
Figure 6.29. Vertex List for the Icosahedron.
A model for the icosahedron is shown in Figure 6.30. The face list for the icosahedron can be read directly off
of it. (Question: what is the normal vector to face #8?)
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 15

11
10

9
15

7
18

13
4

9
8

0
4

6
2

8
3

4
4
Figure 6.30. Model for the icosahedron.
Some people prefer to adjust the model for the icosahedron slightly into the form shown in Figure 6.31. This
makes it clearer that an icosahedron is made up of an anti-prism (shown shaded) and two pentagonal pyramids
on its top and bottom.
5 faces for
pyramidal cap
10 faces make
an antiprism
5 faces for
pyramidal base

Figure 6.31. An icosahedron is an anti-prism with a cap and base.

The Dodecahedron
The dodecahedron is dual to the icosahedron, so all the information needed to build lists for the dodecahedron
is buried in the lists for the icosahedron. But it is convenient to see the model of the dodecahedron laid out, as
in Figure 6.32.

2
1

12
12
13

0
0 4

13
18 13
11

15
8

10
9

6 7
7
1

9
19

2
8

14
19
19

Figure 6.32. Model for the dodecahedron.

Using duality again, we know vertex k of the dodecahedron lies at the center of face k of the icosahedron: just
average the three vertices of face k. All of the vertices of the dodecahedron are easily calculated in this fashion.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 16

Practice Exercises
6.3.9. Icosahedral Distances. What is the radial distance of each vertex of the icosahedron above from the
origin?
6.3.10. Vertex list for the dodecahedron. Build the vertex list and normal list for the dodecahedron.
6.3.3. Other interesting Polyhedra.
There are endless varieties of polyhedra (see for instance [Wenn71] and [Coxe63 polytopes]), but one class is
particularly interesting. Whereas each Platonic solid has the same type of n-gon for all of its faces, the Archimedian (also semi-regular) solids have more than one kind of face, although they are still regular polygons. In addition it is required that every vertex is surrounded by the same collection of polygons in the same
order.
For instance, the truncated cube, shown in Figure 6.33a, has 8-gons and 3-gons for faces, and around each
vertex one finds one triangle and two 8-gons. This is summarized by associating the symbol 388 with this
solid.
a).
b).
c).
1
A

Figure 6.33. The Truncated Cube.

The truncated cube is formed by slicing off each corner of a cube in just the right fashion. The model for the
truncated cube, shown in Figure 6.33b, is based on that of the cube. Each edge of the cube is divided into three
parts, the middle part of length A = 1/(1 + 2) (Figure 6.33c), and the middle portion of each edge is joined
to its neighbors. Thus if an edge of the cube has endpoints C and D, two new vertices V, and W are formed as
the affine combinations

1+ A
1 A
C+
D
2
2
1+ A
1 A
W=
C+
D
2
2
V=

(6.5)

Based on this it is straightforward to build vertex and face lists for the truncated cube (see the exercises and
Case Study 6.3).
Given the constraints that faces must be regular polygons, and that they must occur in the same arrangement
about each vertex, there are only 13 possible Archimedian solids discussed further in Case Study 6.10. Archimedean solids still enjoy enough symmetry that the normal vector to each face is found using the center of
the face.
One Archimedian solid of particular interest is the truncated icosahedron 562 shown in Figure 6.34, which
consists of regular hexagons and pentagons. The pattern is familiar from soccer balls used around the world.
More recently this shape has been named the Buckyball after Buckminster Fuller because of his interest in
geodesic structures similar to this. Crystallographers have recently discovered that 60 atoms of carbon can be

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 17

arranged at the vertices of the truncated icosahedron, producing a new kind of carbon molecule that is neither
graphite nor diamond. The material has many remarkable properties, such as high-temperature stability and
superconductivity [Brow90 Sci. Amer., also see Arthur hebard, Science Watch 1991], and has acquired the
name Fullerine.

Figure 6.34. The Bucky Ball.

The Bucky ball is interesting to model and view on a graphics display. To build its vertex and face lists, draw
the model of the icosahedron given in Figure 6.32 and divide each edge into 3 equal parts. This produces two
new vertices along each edge, whose positions are easily calculated. Number the 60 new vertices according to
taste to build the vertex list of a Bucky ball. Figure 6.35 shows a partial model of the icosahedron with the new
vertices connected by edges. Note that each old face of the icosahedron becomes a hexagon, and that each old
vertex of the icosahedron has been snipped off to form a pentagonal face. Building the face list is just a
matter of listing what is seen in the model. Case Study 6.10 explores the Archimedean solids further.
16

10
9
8

5
1

11
7

2
6

Figure 6.35. Building a Bucky Ball.

Geodesic Domes.
Few things are harder to put up with than a good example.
Mark Twain
Although Buckminster Fuller was a pioneer along many lines, he is best known for introducing geodesic
domes [fuller73]. They form an interesting class of polyhedra having many useful properties. In particular, a
geodesic dome constructed out of actual materials exhibits extraordinary strength for its weight.
There are many forms a geodesic dome can take, and they all approximate a sphere by an arrangement of
faces, usually triangular in shape. Once the sphere has been approximated by such faces, the bottom half is
removed, leaving the familiar dome shape. An example is shown in Figure 6.36 based on the icosahedron.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 18

Figure 6.36. A geodesic dome.

To determine the faces, each edge of the icosahedron is subdivided into 3F equal parts, where Fuller called F
the frequency of the dome. In the example F = 3, so each icosahedral edge is trisected to form 9 smaller triangular faces (see Figure 6.37a).
a).

b).

P18
P81

w18

V8
w81

Figure 6.37. Building new vertices for the dome.

These faces do not lie in the plane of the original faces, however; first the new vertices are projected outward onto the surrounding sphere. Figure 6.37b presents a cross sectional view of the projection process. For
example, the edge from V1 to V8 is subdivided to produce the two points, W18 and W81. Vertex W18 is given by:

W18 =

2
1
V1 + V8 .
3
3

(6.6)

(What is W81?). Projecting W18 onto the enclosing sphere of radius R is simply a scaling:

P18 = R

w 18
| w18 |

(6.7)

where we write w18 for the position vector associated with point W18. The old and new vertices are connected
by straight lines to produce the nine triangular faces for each face of the icosahedron. (Why isnt this a new
Platonic solid?) What are the values for E, F, and V for this polyhedron? Much more can be found on geodesic
domes in the references, particularly [full75] and [kapp91].
Practice Exercises.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 19

6.3.11. Lists for the truncated icosahedron. Write the vertex, normal, and face lists for the truncated icosahedron described above.
6.3.12. Lists for a Bucky Ball. Create the vertex, normal, and face lists for a Bucky ball. As computing 60
vertices is tedious, it is perhaps easiest to write a small routine to form each new vertex using the vertex list of
the icosahedron of Figure 6.29.
6.3.13. Build lists for the geodesic dome. Construct vertex, normal, and face lists for a frequency 3 geodesic
dome as described above.

6.4. Extruded Shapes

A large class of shapes can be generated by extruding or sweeping a 2D shape through space. The prism
shown in Figure 6.19 is an example of sweeping linearly, that is in a straight line. As we shall see, the tetrahedron and octahedron of Figure 6.22 are also examples of extruding a shape through space in a certain way.
And surfaces of revolution can also be approximated by extrusion of a polygon, once we slightly broaden the
definition of extrusion.
In this section we examine some ways to generate meshes by sweeping polygons in discrete steps. In Section
6.5 we develop similar tools for building meshes that attempt to approximate smoothly swept shapes.
6.4.1. Creating Prisms.
We begin with the prism, which is formed by sweeping a polygon in a straight line. Figure 6.38 shows a prism
based on a polygon P lying in the xy-plane. P is swept through a distance H along the z-axis, forming the ARROW prism shown in Figure 6.38b. (More generally, the sweep could be along a vector d, as in Figure 6.19.)
As P is swept along, an edge in the z-direction is created for each vertex of P. This creates another 7 vertices,
so the prism has 14 vertices in all. They appear in pairs: if (xi,yi, 0) is one of the vertices of the base P, then the
prism also contains the vertex (xi,yi,H).
a). polygon base:

b). P swept in z-direction

"cap"

z
y

y
ARROW

P
x
x

c). Model for ARROW prism

3
5
#7

8 9

10 11

Figure 6.38. An example prism.

What face list describes ARROW? The prism is unfolded into the model shown in Figure 6.38c to expose its
nine faces as seen from the outside. There are seven rectangular sides plus the bottom base P and the top cap.
Face 2, for instance, is defined by vertices 2, 9, 10, and 3.
Because the prism has flat faces we associate the same normal vector with every vertex of a face: the normal
vector to the face itself.
Building a mesh for the prism.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 20

We want a tool to make a mesh for the prism based on an arbitrary polygon. Suppose the prism's base is a
polygon with N vertices (xi, yi). We number the vertices of the base 0, . . . , N-1 and those of the cap N, . . .,
2N -1, so that an edge joins vertices i and i + N, as in the example. The vertex list is then easily constructed to
contain the points (xi, yi, 0) and (xi, yi, H), for i = 0, 1, ..., N-1.
The face list is also straightforward to construct. We first make the side faces or walls, and then add the
cap and base. For the j-th wall (j = 0,...,N-1) we create a face with the four vertices having indices j, j + N,
next(j)+ N, and next(j) where next(j) is j+1 except if j equals N-1, whereupon it is 0. This takes care of the
wrap-around from the (N-1)st to the 0-th vertex. next(j) is given by:
next(j) = (j+1) modulo N

(6.8)

or in terms of program code: next = (j < (N-1)) ? (j + 1) : 0. Each face is inserted in the face
list as it is created. The normal vector to each face is easily found using the Newell method described earlier.
We then create the base and cap faces and insert them in the face list. Case Study 6.3 provides more details for
building mesh models of prisms.
6.4.2. Arrays of Extruded Prisms - brick laying.
Some rendering tools, like OpenGL, can reliably draw only convex polygons. They might fail, for instance, to
draw the arrow of Figure 6.38 correctly. If this is so the polygon can be decomposed (tesselated) into a set of
convex polygons, and each one can be extruded. Figure 6.39 shows some examples, including a few extruded
letters of the alphabet.

Figure 6.39. Extruded objects based on a collection of convex prisms.

Prisms like this are composed of an array of convex prisms. Some of the component prisms abut one another
and therefore share all or parts of some walls. Because vertex positions are being computed at a high precision,
the crease where two walls adjoin will usually be invisible.
For this family of shapes we need a method that builds a mesh out of an array of prisms, say
void Mesh:: makePrismArray(...)
which would take as its arguments a suitable list of (convex) base polygons (assumed to lie in the xy-plane),
and perhaps a vector d that describes the direction and amount of extrusion. The vertex list would contain the
vertices of the cap and base polygons for each prism, and the individual walls, base, and cap of each prism
would be stored in the face list. Drawing such a mesh would involve some wasted effort, since walls that abut
would be drawn (twice), even though they are ultimately invisible.
Special case: Extruded Quad-Strips.
A simpler but very interesting family of such prisms can be built and manipulated more efficiently. These are
prisms for which the base polygon can be represented by a quad-strip. The quad-strip is an array of quadrilaterals connected in a chain, like laying bricks in a row, such that neighboring faces coincide completely, as
shown in Figure 6.40a. Recall from Figure 2.37 that the quad-strip is an OpenGL geometric primitive. A quadstrip is described by a sequence of vertices

Figure 6.40. Quad-strips and prisms built upon quad-strips.

quad-strip = {p0, p1, p2, ...., pM-1}

Hill - ECE660:Computer Graphics Chapter 6

(6.9)

10/23/99

page 21

The vertices are understood to be taken in pairs, with the odd ones forming one edge of the quad-strip, and
the even ones forming the other edge. Not every polygon can be represented as a quad-strip. (Which of the
polygons in Figure 6.39 are not quad-strips? What block letters of the alphabet can be drawn as quad-strips?)
When a mesh is formed as an extruded quad-strip only 2M vertices are placed in the vertex list, and only the
outside walls are included in the face list. There are 2M - 2 faces in all. (Why?) Thus no redundant walls are
drawn when the mesh is rendered. A method for creating a mesh for an extruded quad-strip would take an array of 2D points and an extrusion vector as its parameters:
void Mesh:: makeExtrudedQuadStrip(Point2 p[], int numPts, Vector3 d);
Figure 6.41 shows some examples of interesting extruded quad-strips. Case Study 6.4 considers how to make
such meshes in more detail.

Figure 6.41. Extruded quad-strips - arches.

6.4.3. Extrusions with a twist.
So far an extrusion just shifts the base polygon to a new position to define the cap polygon. It is easy to generalize on this in a way that produces a much broader family of shapes; create the cap polygon as an enlarged or
shrunk, and possibly rotated, version of the base polygon. Specifically, if the base polygon is P, with vertices
{p0, p1, ..., pN-1}, the cap polygon has vertices
P = { Mp0, Mp1, ..., MpN-1}

(6.10)

where M is some 4 by 4 matrix representing an affine transformation. Figure 6.42 shows some examples. Parts
a and b show pyramids, or tapered cylinders (also truncated cones), where the cap is a smaller version of
the base. The transformation matrix for this is:

Figure 6.42. Pyramids and twisted prisms.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 22

0.7
0
M =
0
0

0
0.7
0
0

0 0
0 0
1 H
0 1

based simply on a scaling factor of 0.7 and a translation by H along z. Part c shows a cylinder where the cap
has been rotated through an angle about the z-axis before translation, using the matrix:

cos( )
sin( )
M =
0
0

sin( )
cos( )
0
0

0
0
1
0

H

1
0
0

And part d shows in cross section how cap P can be rotated arbitrarily before it is translated to the desired
position.
Prisms such as these are just as easy to create as those that use a simple translation for M: the face list is identical to the original; only the vertex positions (and the values for the normal vectors) are altered.
Practice exercises.
6.4.1. The tapered cylinder. Describe in detail how to make vertex, normal, and face lists for a tapered cylinder having regular pentagons for its base and cap, where the cap is one-half as large as the base.
6.4.2. The tetrahedron as a tapered cylinder. Describe how to model a tetrahedron as a tapered cylinder
with a triangular base. Is this an efficient way to obtain a mesh for a tetrahedron?
6.4.3. An anti-prism. Discuss how to model the anti-prism shown in Figure 6.20b. Can it be modeled as a
certain kind of extrusion?
6.4.4. Building Segmented Extrusions - Tubes and Snakes.
Another rich set of objects can be modeled by employing a sequence of extrusions, each with its own transformation, and laying them end-to-end to form a tube. Figure 6.43a shows a tube made by extruding a
square P three times, in different directions with different tapers and twists. The first segment has end polygons M0P and M1P, where the initial matrix M0 positions and orients the starting end of the tube. The second
segment has end polygons M1P and M2P, etc. We shall call the various transformed squares the waists of the
tube. In this example the vertex list of the mesh contains the 16 vertices M0p0, M0 p1, M0 p2, M0 p3, M1p0, M1p1,
M1p2, M1p3, ..., M3p0, M3p1, M3p2, M3p3. Figure 6.43b shows a snake, so called because the matrices Mi cause
the tube to grow and shrink to represent the body and head of a snake.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 23

Figure 6.43*. A tube made from successive extrusions of a polygon.

Designing tubes based on 3D curves.
How do we design interesting and useful tubes and snakes? We could choose the individual matrices Mi by
hand, but this is awkward at best. It is much easier to think of the tube as wrapped around a curve which we
shall call the spine of the tube that undulates through space in some organized fashion5. We shall represent the
curve parametrically as C(t). For example, the helix (recall Section 3.8) shown in Figure 6.44a has the parametric representation
stereo pair of helix
Figure 6.44. a helix - stereo pair.
C(t) = (cos(t), sin(t), bt)

(6.11)

for some constant b.

To form the various waist polygons of the tube we sample C(t) at a set of t-values, {t0, t1, ...}, and build a transformed polygon in the plane perpendicular to the curve at each point C(ti) , as suggested in Figure 6.45. It is
convenient to think of erecting a local coordinate system at each chosen point along the spine: the local zaxis points along the curve, and the local x and y-axes point in directions normal to the z-axis (and normal
to each other). The waist polygon is set to lie in the local xy-plane. All we need is a straightforward way to
determine the vertices of each waist polygon.

Figure 6.45. Constructing local coordinate systems along the spine curve.
It is most convenient to let the curve C(t) itself determine the local coordinate systems. A method well-known
in differential geometry creates the Frenet frame at each point along the spine [gray 93]. At each value ti of
interest a vector T(ti) that is tangent to the curve is computed. Then two vectors, N(ti) and B(ti), which are perpendicular to T(ti) and to each other, are computed. These three vectors constitute the Frenet frame at ti.
Once the Frenet frame is computed it is easy to find the transformation matrix M that transforms the base
polygon of the tube to its position and orientation in this frame. It is the transformation that carries the world

The VRML 2.0 modeling language includes an extrusion node that works in a similar fashion, allowing the
designer to define a spine along which the polygons are placed, each with its own transformation.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 24

coordinate system into this new coordinate system. (The reasoning is very similar to that used in Exercise 5.6.1
on transforming the camera coordinate system into the world coordinate system.) The matrix Mi must carry i, j,
and k into N(ti), B(ti), T(ti), respectively, and must carry the origin of the world into the spine point C(t). Thus
the matrix has columns consisting directly of N(ti), B(ti), T(ti), and C(ti) expressed in homogeneous coordinates:

M i = N(t i )| B(t i )| T(t i )| C(t i )

(6.12)

Forming the Frenet frame.

The Frenet frame at each point along a curve depends on how the curve twists and undulates. It is derived from
certain derivatives of C(t), and so it is easy to form if these derivatives can be caclulated.
Specifically, if the formula that we have for C(t) is differentiable, we can take its derivative and form the tan (t ) . (If C(t) has components C (t), C (t), and C (t) this derivative is
gent vector to the curve at each point, C
x
y
z
(t ) = ( C
(t ), C
(t ), C
(t )) . This vector points in the direction the curve is headed at each value
simply C
x
y
z

of t, that is in the direction of the tangent to the curve. We normalize it to unit length to obtain the unit tangent vector at t. For example, the helix of Equation 6.11 has the unit tangent vector given by

T(t ) =

1
1 + b2

( sin(t ), cos(t ), b)

(6.13)

This tangent is shown for various values of t in Figure 6.46a.

Figure 6.46*. a). Tangents to the helix. B). Frenet frame at various values of t, for the helix.
If we form the cross product of this with any non-collinear vector we must obtain a vector perpendicular
to T(t) and therefore perpendicular to the spine of the curve. (Why?) A particularly good choice is the
(t ) . So we form C
(t ) C
(t ) , and since it will be used
acceleration, based on the second derivative, C
for an axis of the coordinate system, we normalize it, to obtain the unit binormal vector as:

B(t ) =

(t ) C
(t )
C
(t ) C
(t )
C

(6.14)

We then obtain a vector perpendicular to both T(t) and B(t) by using the cross product again:

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 25

N(t ) = B (t ) T(t )

(6.15)

Convince yourself that these three vectors are mutually perpendicular and have unit length, and thus
constitute a local coordinate system at C(t) (known as a Frenet frame). For the helix example these
vectors are given by:

B(t ) =

(b sin(t ), b cos(t ),1)

1 + b2
N(t ) = ( cos(t ), sin(t ),0)

(6.16)

Figure 6.46b shows the Frenet frame at various values of t along the helix.
Aside: Finding the Frenet frame Numerically.
If the formula for C(t) is complicated it may be awkward to form its successive derivatives in closed
form, such that formulas for T(t), B(t), and N(t) can be hard-wired into a program. As an alternative, it is
possible to approximate the derivatives numerically using:

C ( t + ) C (t )
2
C ( t ) 2 C (t ) + C (t + )
(t ) =
C

2
(t ) =
C

(6.17)

This computation will usually produce acceptable directions for T(t), B(t), and N(t), although the user
should beware that numerical differentiation is an inherently unstable process [burden85].
Figure 4.47 shows the result of wrapping a decagon about the helix in this way. The helix was sampled at
30 points, a Frenet frame was constructed at each point, and the decagon was erected in the new frame.

Figure 6.47. A tube wrapped along a helix.

Figure 4.48 shows other interesting examples, based on the toroidal spiral (which we first saw in Section
3.8.) [gray93, p.212]. (The edges of the individual faces are drawn to clarify how the tube twists as it proceeds. Drawing the edges of a mesh is considered in Case Study 6.7.) A toroidal spiral is formed when a spiral is wrapped about a torus (try to envision the underlying invisible torus here), and it is given by

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 26

Figure 6.48. Tubes based on toroidal spirals. (file: torusKnot.bmp, file: torusKnot7.bmp)

C(t ) = (( a + b cos(qt )) cos( pt ),( a + b cos( qt ))sin( pt ), c sin( qt ))

(6.18)

for some choice of constants a, b, p, and q. For part a the parameters p and q were chosen as 2 and 5, and
for part b they are chosen to be 1 and 7.
Figure 6.49 shows a sea shell, formed by wrapping a tube with a growing radius about a helix. To accomplish this, the matrix of Equation 6.12 was multiplied by a scaling matrix, where the scale factors also depend
on t:

Figure 6.49. A sea shell.

g( t )
0
M' = M
0
0

0
g( t )
0
0

0
0
1
0

0
0
0
1

Here g(t) = t. It is also possible to add a rotation to the matrix, so that the tube appears to twist more vigorously as one looks along the spine.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 27

One of the problems with using Frenet frames for sweeping curves is that the local frame sometimes
twists in such a way as to introduce undesired knots in the surface. Recent work, such as [wang97],
finds alternatives to the Frenet frame that produce less twisting and therefore more graceful surfaces.
Practice Exercises.
2

( t ) ( C
(t ) C
( t )) C
(t ) / C
( t ) , so it points in
6.4.4. What is N(t)? Show that N(t) is parallel to C

the direction of the acceleration when the velocity and acceleration at t are perpendicular.
6.4.5. The Frame for the helix. Consider the circular helix treated in the example above. Show that the
formulas above for the unit tangent, binormal, and normal vectors are correct. Also show that these are
unit length and mutually perpendicular. Visualize how this local coordinate system orients itself as you
move along the curve.
Figure 6.50 shows additional examples. Part (a) shows a hexagon wrapped about an elliptical spine to form a
kind of elliptical torus, and part (b) shows segments arranged into a knot along a Lissajous figure given by:

C(t ) = (r cos( Mt + ),0, r sin( Nt ))

(6.19)

with M = 2 , N = 3 , = 0.

Figure 6.50. a). A hexagon wrapped about an elliptical torus. b). A 7-gon wrapped about a Lissajous figure.
Case Study 6.5 examines more details of forming meshes that model tubes based on a parametric curve.
6.4.5. Discretely Swept Surfaces of Revolution.
The tubes above use affine transformations to fashion a new coordinate system at each spine point. If we use
pure rotations for the affine transformations, and place all spine points at the origin, a rich set of polyhedral
shapes emerges. Figure 6.51 shows an example, where a base polygon - now called the profile - is initially
positioned 3 units out along the x-axis, and then is successively rotated in steps about the y-axis to form an
approximation of a torus.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 28

Figure 6.51. Rotational sweeping in discrete steps.

This is equivalent to circularly sweeping a shape about an axis, and the resulting shape is often called a surface of revolution. We examine true surfaces of revolution in Section 6.5; here we are forming only a discrete
approximation to them since we are sweeping in discrete steps.
Figure 6.52 shows an example that produces pictures of a martini glass. The profile here is not a closed polygon but a simple polyline based on points Pj = (xj, yj, 0). If we choose to place this polyline at K equispaced
angles about the y-axis, we set the transformations to have matrices:

Figure 6.52. Approximating a martini glass with a discretely swept polyline. a). the profile, b). the swept surface.

cos( )
0
~
M =
sin( )
0
i

0 sin( i )
1
0
0 cos( i )
0
0

0
0
0
1

where i = 2i/K, i = 0, 1, ..., K-1. Note there is no translation involved. This transformation is simple enough
that we can write the positions of the vertices directly. The rotation sets the points of the i-th waist polyline at:
(xj cos(i), yj, xj sin(i))

(6.20)

Building meshes that model surfaces of revolution is treated further in Case Study 6.6.

6.5. Mesh Approximations to Smooth objects.

So far we have built meshes to represent polyhedra, where each shape is a collection of flat polygonal faces.
They tend to be data-intensive, specified by listing the vertices of each face individually. Now we want to
build meshes that attempt to approximate inherently smooth shapes like a sphere or torus. These shapes are
normally defined by formulas rather than data. We also want to arrange matters so that these meshes can be
smoothly shaded: even though they are represented by a collection of flat faces as before, the proper algorithm
(Gouraud shading) draws them with smooth gradations in shading, and the individual faces are invisible (recall
Figure 6.1). All this requires is that we find the proper normal vector at each vertex of each face. Specifically,

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 29

we compute the normal vector to the underlying smooth surface. We discuss the Gouraud shading algorithm in
Chapter 11.
The basic approach for each type of surface is to polygonalize (also called tesselate) it into a collection of flat
faces. If the faces are small enough and there is a graceful change in direction from one face to the next, the
resulting mesh will provide a good approximation to the underlying surface. The faces have vertices that are
found by evaluating the surfaces parametric representation at discrete points. A mesh is created by building a
vertex list and face list in the usual way, except here the vertices are computed from formulas. The same is
true for the vertex normal vectors: they are computed by evaluating formulas for the normal to the surface at
discrete points.
6.5.1. Representations for Surfaces.
To set the stage, recall that in Section 4.5.5 we examined the planar patch, given parametrically by
P(u, v) = C + au + bv

(6.21)

where C is a point, and a and b are vectors. The range of values for the parameters u and v is usually restricted to [0, 1], in which case the patch is a parallelogram in 3D with corner vertices C, C + a, C + b,
and C + a + b, (recall Figure 4.31).
Here we enlarge our interests to nonlinear forms, to represent more general surface shapes. We introduce
three functions X(), Y(), and Z() so that the surface has parametric representation in point form
P(u, v) = (X(u, v), Y(u, v), Z(u, v))

(6.22)

with u and v restricted to suitable intervals. Different surfaces are characterized by different functions, X,
Y, and Z. The notion is that the surface is at (X(0, 0), Y(0, 0), Z(0, 0)) when both u and v are zero, at
(X(1, 0), Y(1, 0), Z(1, 0)) when u = 1 and v = 0, and so on. Keep in mind that two parameters are required when representing a surface, whereas a curve in 3D requires only one. Letting u vary while
keeping v constant generates a curve called a v-contour. Similarly, letting v vary while holding u constant produces a u-contour. (Look ahead to Figure 6.58 to see examples u- and v-contours.)
The implicit form of a surface.
Although we are mainly concerned with parametric representations of different surfaces, it will prove
useful to keep track of an alternative way to describe a surface, through its implicit form. Recall from
Section 3.8 that a curve in 2D has an implicit form F(x, y) which must evaluate to 0 for all points (x, y)
that lie on the curve, and for only those. For surfaces in 3D a similar functions F(x, y, z) exists that
evaluates to 0 if and only if the point (x, y, z) is on the surface. The surface therefore has an implicit
equation given by
F(x, y, z) = 0

(6.23)

that is satisfied for all points on the surface, and only those. The equation constrains the way that values
of x, y, and z must be related to confine the point (x, y, z) to the surface in question. For example, recall
(from Chapter 4) that the plane that passes through point B and has normal vector n is described by the
equation nx x + ny y + nz z = D (where D = nB), so the implicit form for this plane is F(x, y, z) = nx x +
ny y + nz z - D. Sometimes it is more convenient to think of F as a function of a point P, rather than a
function of three variables x, y, and z, and we write F(P) = 0 to describe all points that lie on the surface.
For the example of the plane here, we would define F(P) = n(P - B) and say that P lies in the plane if
and only if F(P) = n(P-B) is zero. If we wish to work with coordinate frames (recall Section 4.5) so
~ ~ ~
~
~
T
that P is the 4-tuple P = ( x, y, z,1) , the implicit form for a plane is even simpler: F ( P ) = n
P,
~
where n = ( nx , ny , nz , D) captures both the normal vector and the value D.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 30

It is not always easy to find the function F(x, y, z) or F(P) from a given parametric form (nor can you
always find a parametric form when given F(x, y, z)). But if both a parametric form and an implicit form
are available it is simple to determine whether they describe the same surface. Simply substitute X(u, v),
Y(u, v), and Z(u, v) for x, y, and z, respectively, in F(x, y, z) and check that F is 0 for all values of u and v
of interest.
For some surfaces like a sphere that enclose a portion of space it is meaningful to define an inside region and an outside region. Other surfaces like a plane clearly divide 3D space into two regions, but
one must refer to the context of the application to tell which half-space is the inside and which the outside. There are also many surfaces, such as a strip of ribbon candy, for which it makes little sense to
name an inside and an outside.
When it is meaningful to designate an inside and outside to a surface, the implicit form F(x, y, z) of a
surface is also called its inside outside function. We then say that a point (x, y, z) is
inside the surface if:
on the surface if:
outside the surface if:

F(x, y, z) < 0
F(x, y, z) = 0
F(x, y, z) > 0

(6.24)

This provides a quick and simple test for the disposition of a given point (x', y', z') relative to the surface:
Just evaluate F(x', y', z') and test whether it is positive, negative, or zero. This is seen to be useful in hidden line and hidden surface removal algorithms in Chapter 14, and in Chapter 15 it is used in ray tracing
algorithms. There has also been vigorous recent activity in rendering surfaces directly from their implicit
forms: see [bloomenthal, 97].
6.5.2. The Normal Vector to a Surface.
As described earlier, we need to determine the direction of the normal vector to a surface at any desired
point. Here we present one way based on the parametric expression, and one based on the implicit form,
of the surface. As each surface type is examined later, we find the suitable expressions for its normal
vector at any point.
The normal direction to a surface can be defined at a point, P(u0, v0), on the surface by considering a
very small region of the surface around P(u0, v0). If the region is small enough and the surface varies
smoothly in the vicinity, the region will be essentially flat. Thus it behaves locally like a tiny planar
patch and has a well-defined normal direction. Figure 6.53 shows a surface patch with the normal vector
drawn at various points. The direction of the normal vector is seen to be different at different points on
the surface.

Figure 6.53*. The normal vector to a surface.

We use the name n(u, v) for the normal at (u, v). We now examine how it can be calculated.
The Normal Vector for a Surface Given Parametrically.
Not surprisingly, n(u0, v0) takes the form of a cross product between two vectors that lie in the tiny planar
patch near (u0, v0). Being a cross product it is guaranteed to be perpendicular to both vectors. The two
vectors in the plane (indicated as tu and tv in the figure) are certain tangent vectors. Calculus texts show
that they are simply related to partial derivatives of p(u, v) (the vector from the origin to the surface
point P(u, v)6), evaluated at the point in question [thomas53]. An expression for the normal vector is
therefore

Since p(u, v) is simply the difference P(u, v) - (0,0,0), the derivative of p() is the same as that of P().

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 31

n(u0 , v0 ) =

p p
u v

(6.25)
u = u0 ,v = v0

where the vertical bar | indicates that the derivatives are evaluated at u = u0, v = v0. Formed this way,
n(u0, v0) is not automatically a unit length vector, but it can be normalized if desired.
Example 6.5.1: Does this work for a plane? Consider the plane given parametrically by P(u, v) = C +
au + bv. The partial derivative of this with respect to u is just a, and that with respect to v is b. Thus according to Equation 6.19, n(u, v) = a b , which we recognize as the correct result.
More generally, the partial derivatives of p(u, v) exist whenever the surface is smooth enough. Most
of the surfaces of interest to us in modeling scenes have the necessary smoothness and have simple
enough mathematical expressions so that finding the required derivatives is not difficult. Because p(u, v)
= X(u, v)i + Y(u, v)j + Z(u, v)k, the derivative of a vector is just the vector of the individual derivatives:

p(u, v )
X (u, v ) Y (u, v ) Z (u, v )
,
,
=
u
u
u
u

(6.26)

We apply these formulas directly to each surface type we examine later.

The Normal Vector for a Surface given Implicitly.
An alternative expression is used if the surface is given by an implicit form, F(x, y, z) = 0. The normal
direction at the surface point, (x, y, z), is found using the gradient,
F , of F, which is given by [thomas 53]:

n( x 0 , y 0 , z 0 ) = F

x = x0 , y = y 0 , z = z 0

F , F , F
x y z

(6.27)
x = x0 , y = y 0 , z = z 0

where each partial derivative is evaluated at the desired point, (x0, y0, z0). If the point (x0, y0, z0) for the
surface in question corresponds to the point P(u0,v0) of the parametric form, then n(x0, y0, z0) has the same
direction as n(u0,v0) in Equation 6.19, but it may have a different length. Again, it can be normalized if
desired.
Example 6.5.2. The plane again: Consider once again the plane with normal n that passes through point
A, given implicitly by F ( x , y , z ) = n (( x , y , z ) A) = 0 , or nx x + n y y + nz z n A = 0 . This has
gradient F = n as expected.
Note that the gradient-based form gives the normal vector as a function of x, y, and z, rather than of u
and v. Sometimes for a surface we know both the inside-outside function, F(x, y, z), and the parametric
form, p(u, v) = X(u, v)i + Y(u, v)j + Z(u, v)k. In such cases it may be easiest to find the parametric form,
n(u, v), of the normal at (u, v) by a two-step method: (1) Use Equation 6.21 to get the normal at (x, y, z)
in terms of x, y, and z, and then (2) substitute the known functions X(u, v) for x, Y(u, v) for y, and Z(u, v)
for z. Some of the later examples illustrate this method.
6.5.3. The Effect of an Affine Transformation.
We shall need on occasion to work with the implicit and parametric forms of a surface after the surface
has been subjected to an affine transformation. We will also want to know how the normal to the surface
is affected by the transformation.
Suppose the transformation is represented by 4 by 4 matrix M, and that the original surface has implicit
~
form (in terms of points in homogeneous coordinates) F ( P ) and parametric

form P (u, v ) = ( X (u, v), Y (u, v ), Z * (u, v ),1) . Then it is clear that the transformed surface has para-

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 32

metric form MP (u, v) (why?). It is also easy to show (see the exercises) that the transformed surface

has implicit form F ( P ) given by:

~
~
F ( P ) = F ( M 1 P )
Further, if the original surface has normal vector n(u, v) then the transformed surface has:

normal vector: M T n(u, v)

~ P with
For example, suppose we transform the plane examined above, given by F ( P ) = n

~
~
n~ = (nx , ny , nz , D) . The transformed plane has implicit form F ( P ) = n~ ( M 1P ) . This can be writ~
T ~
ten (see the exercises) as ( M n
) P , so the normal vector of the transformed plane involves the in-

verse transpose of the matrix, consistent with the claimed form for the normal to a general surface.
Practice Exercises.
6.5.1. The implicit form of a transformed surface. Suppose all points on a surface satisfy F(P) =
~
~
~
~
~
0, and that M transforms P into Q ; i.e. Q = MP . Then argue that any point Q on the transformed
1

surface comes from a point M Q , and those points all satisfy F ( M Q ) = 0 . Show that this

proves that the implicit form for the transformed surface is F' (Q ) = F ( M Q) .
T
6.5.2. How are normal vectors affected? Let n = (nx, ny, nz, 0) be the normal at P and let v be any
vector tangent to the surface at P. Then n must be perpendicular to v and we can write n v = 0.
a). Show that the dot product can be written as a matrix product: nTv = 0 (see Appendix 2).
b). Show that this is still 0 when the matrix product M-1M is inserted: nTM-1Mv = 0.
c). Show that this can be rewritten as (M-Tn)(Mv)=0, so M-Tn is perpendicular to (Mv).
Now since the tangent v transforms to Mv, which is tangent to the transformed surface, show that
this says M-Tn must be normal to the transformed surface, which we wished to show.
d). The normal to a surface is also given by the gradient of the implicit form, so the normal to the
transformed surface at point P must be the gradient of F(M-1P). Show (by the chain rule of calculus)
that the gradient of this function is M-T multiplied onto the gradient of F().
6.5.3. The tangent plane to a transformed surface. To find how normal vectors are transformed
we can also find how the tangent plane to a surface is mapped to the tangent plane on the transformed surface. Suppose the tangent plane to the original surface at point P has parametric representation P + au + bv, where a and b are two vectors lying in the plane. The normal to the surface is
therefore n = a b .
a). Show that the parametric representation of the transformed plane is MP + Mau + Mbv, and that
this plane has normal n' = ( Ma ) ( Mb) .
b). Referring to Appendix 2, show the following identity:

( Ma ) ( Mb) = (det M ) M T (a b)

This relates the cross product of transformed vectors to the cross product of the vectors themselves. c). Show
that n is therefore parallel to M-Tn.
6.5.4. Three generic shapes: the sphere, cylinder, and cone.
We begin with three classic objects, generic versions of the sphere, cylinder, and cone. We develop
the implicit form and parametric form for each of these, and see how one might make meshes to approximate them. We also derive formulas for the normal direction at each point on these objects. Note
that we have already used OpenGL functions in Chapter 5 to draw these shapes. Adding our own tools
has several advantages, however: a). We have much more control over the detailed nature of the shape

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 33

being created; b). We have the object as an actual mesh that can be operated upon by methods of the
Mesh class.
The Generic Sphere.
We call the sphere of unit radius centered at the origin the generic sphere (see Figure 6.54a). It forms
the basis for all other sphere-like shapes we use. It has the familiar implicit form
F(x, y, z) = x2 + y2 + z2 - 1

(6.28)

In the alternate notation F(P) we obtain the more elegant F(P) = |P|2 - 1. (What would these forms be if
the sphere has radius R?)
A parametric description of this sphere comes immediately from the basic description of a point in
spherical coordinates (see Appendix 2). We choose to let u correspond to azimuth and v correspond to
latitude. Then any point P = (x, y, z) on the sphere has the representation (cos(v) cos(u), cos(v) sin(u),
sin(v)) in spherical coordinates (see Figure 6.54b). We let u vary over (0, 2) and v vary over (-/2, /2)
to cover all such points. A parametric form for the sphere is therefore
a). generic sphere b). parametrized by azimuth, latitude. c). pars & merids
Figure 6.54. a). The generic sphere, b) a parametric form. c). parallels and meridians
P(u, v) = (cos(v) cos(u), cos(v) sin(u), sin(v))

(6.29)

Its easy to check that this is consistent with the implicit form: substitute terms of Equation 6.29 into corresponding terms of Equation 6.28 and see that zero is obtained for any value of u and v.
(Question: What is the corresponding parametric form if the sphere instead has radius R and is centered
at (a, b, c)?)
For geographical reasons certain contours along a sphere are given common names. For this parametrization u-contours are called meridians, and v-contours are known as parallels, as suggested in Figure
6.54c. (Note that this classical definition of spherical coordinates, parallels and meridians causes the
sphere to appear to lie on its side. This is simply a result of how we sketch 3D figures, with the y-axis
pointing up.)
Different parametric forms are possible for a given shape. An alternative parametric form for the sphere
is examined in the exercises.
What is the normal direction n(u, v) of the spheres surface at the point specified by parameters (u, v)?
Intuitively the normal vector is always aimed radially outward, so it must be parallel to the vector
from the origin to the point itself. This is confirmed by Equation 6.21: the gradient is simply 2(x, y, z),
which is proportional to P. Working with the parametric form, Equation 6.19 yields n(u, v) = -cos(v)p(u,
v), so n(u, v) is parallel to p(u, v) as expected. The scale factor -cos(v) will disappear when we normalize
n. We must make sure to use p(u,v) rather than -p(u,v) for the normal, so that it does indeed point radially outward.
The Generic Cylinder.
We adopt as the generic cylinder the cylinder whose axis coincides with the z-axis, has a circular cross
section of radius 1, and extends in z from 0 to 1, as pictured in Figure 6.55a.
a). generic cylinder
b). tapered cylinder

Figure 6.55. The generic cylinder and the tapered cylinder.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 34

It is convenient to view this cylinder as one member of the large family of tapered cylinders, as we did
in Chapter 5. Figure 6.55b shows the generic tapered cylinder, having a small radius of s when z = 1.
The generic cylinder is simply a tapered cylinder with s = 1. Further, the generic cone to be examined
next is simply a tapered cylinder with s = 0. We develop formulas for the tapered cylinder with an arbitrary value of s. These provide formulas for the generic cylinder and cone by setting s to 1 or 0, respectively.
If we consider the tapered cylinder to be a thin hollow shell, its wall is given by the implicit form

F( x, y, z ) = x 2 + y 2 (1 + ( s 1)z )2 for 0 < z < 1

(6.30)

and by the parametric form:

P(u, v ) = ((1 + ( s 1)v ) cos(u),(1 + (s 1)v )sin(u), v)

(6.31)

for appropriate ranges of u and v (which ones?). What are these expressions for the generic cylinder
with s = 1?
When it is important to model the tapered cylinder as a solid object, we add two circular discs at its
ends: a base and a cap. The cap is a circular portion of the plane z = 1, characterized by the inequality x2
+ y2 < s2, or given parametrically by P(u, v) = (v cos(u), v sin(u), 1) for v in [0, s]. (What is the parametric representation of the base?)
The normal vector to the wall of the tapered cylinder is found using Equation 6.27. (Be sure to check
this). It is
n(x, y, z) = (x, y, -(s - 1)(1+ (s - 1)z) )

(6.32)

or in parametric form n(u, v) = (cos(u), sin(u), 1 s). For the generic cylinder the normal is simply
(cos(u), sin(u), 0). This agrees with intuition: the normal is directed radially away from the axis of the
cylinder. For the tapered cylinder it is also directed radially, but shifted by a constant z-component.
(What are the normals to the cap and base?)
The Generic Cone.
We take as the generic cone the cone whose axis coincides with the z-axis, has a circular cross section
of maximum radius 1, and extends in z from 0 to 1, as pictured in Figure 6.56. It is a tapered cylinder
with small radius of s = 0. Thus its wall has implicit form
generic cone

Figure 6.56. The generic cone.

2
2
2
F(x, y, z) = x + y - (1 - z) = 0 for 0 < z < 1

(6.33)

and parametric form P(u, v) = ((1-v) cos(u), (1- v) sin(u), v) for azimuth u in [0, 2] and v in [0, 1]. Using the results for the tapered cylinder again, the normal vector to the wall of the cone is (x, y, 1-z). What
is it parametrically?
For easy reference Figure 6.57 shows the normal vector to the generic surfaces we have discussed.
surface
F( x , y, z )
n(u, v) at p(u, v)
sphere
tapered cylinder

p(u, v)
(cos(u), sin(u), 1 s)

Hill - ECE660:Computer Graphics Chapter 6

(x, y, z)
(x, y, -(s - 1)(1+ (s-1)z))

10/23/99

page 35

cylinder
(cos(u), sin(u), 0)
cone
(cos(u), sin(u), 1)
Figure 6.57. Normal vectors to the generic surfaces.

(x, y, 0)
(x, y, 1 - z)

Practice Exercises.
6.5.4. Alternative representation for the generic sphere. We can associate different geometric quantities with the parameters u and v and obtain a different parametric form for the sphere. We again use parameter u for azimuth, but use v for the height of the point above the xy-plane. All points at height v lie
on a circle of radius
P2(u, v) = (

1 v 2 , so this parametric form is given by:

1 v 2 cos(u), 1 v 2 sin (u), v)

(6.34)

for u in [0, 2] and v in [-1,1]. Show that P2 lies unit distance from the origin for all u and v.
6.5.5. Whats the surface? Let A be a fixed point with position vector a, and P be an arbitrary point
with position vector p. Describe in words and sketch the surface described by: a). pa = 0; b). pa = |a|;
c). |p x a| = |a|; d). p a = p p; e). p a = |a||p|/2.
6.5.6. Finding the normal vector to the generic cylinder and cone. Derive the normal vector for the
generic tapered cylinder and the generic cone in two ways:
a). Using the parametric representation;
b). Using the implicit form, then expressing the result parametrically.
6.5.7. Transformed Spheres. Find the implicit form for a generic sphere that has been scaled in x by 2
and in y by 3, and then rotated 30o about the z-axis.
6.5.5. Forming a Polygonal Mesh for a Curved Surface.
Now we examine how to make a mesh object that approximates a smooth surface such as the sphere,
cylinder, or cone. The process is called polygonalization or tesselation, and it involves replacing the
surface by a collection of triangles and quadrilaterals. The vertices of these polygons lie in the surface itself, and they are joined by straight edges (which usually do not lie in the surface). One proceeds by
choosing a number of values of u and v, and sampling the parametric form for the surface at these
values to obtain a collection of vertices. These vertices are placed in a vertex list. A face list is then created: each face consists of three or four indices pointing to suitable vertices in the vertex list. Associated
with each vertex in a face is the normal vector to the surface. This normal vector is the normal direction
to the true underlying surface at each vertex location. (Note how this contrasts with the normal used
when representing a flat-faced polyhedron: there the vertex of each face is associated with the normal to
the face.)
Figure 6.58 shows how this works for the generic sphere. We think of slicing up the sphere along azimuth lines and latitude lines. Using OpenGL terminology of slices and stacks (see Section 5.6.3), we
choose to slice the sphere into nSlices slices around the equator, and nStacks stacks from the
south pole to the north pole. The figure shows the example of 12 slices and 8 stacks. The larger
nSlices and nStacks are, the better the mesh approximates a true sphere.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 36

Figure 6.58*. a). A mesh approximation to the generic sphere. b). Numbering the vertices.
To make slices we need nSlices values of u between 0 and 2. Usually these are chosen to be equispaced: ui = 2i/nSlices, i = 0, 1, ..., nSlices -1. As for stacks, we put half of them above the
equator and half below. The top and bottom stacks will consist of triangles; all other faces will be quadrilaterals. This requires we define (nStacks + 1) values of latitude: vj = - j/nStacks, j = 0, 1, ...,
nStacks.
The vertex list can now be created. The figure shows how we might number the vertices (the ordering is
a matter of convenience): We put the north pole in pt[0], the bottom points of the top stack into the next
12 vertices, etc. With 12 slices and 8 stacks there will be a total of 98 points (why?)
The normal vector list is also easily created: norm[k] will hold the normal for the sphere at vertex
pt[k]. norm[k] is computed by evaluating the parametric form of n(u,v) at the same (u,v) used for the
points. For the sphere this is particularly easy since norm[k] is the same as pt[k].
For this example the face list will have 96 faces, of which 24 are triangles. We can put the top triangles
in the first 12 faces, the 12 quadrilaterals of the next stack down in the next 12 faces, etc. The first few
faces will contain the data:
number of vertices:
vertex indices:
normal indices:

3
012
012

3
023
023

3
034
034

...
...
...

Note that for all meshes that try to represent smooth shapes the normIndex is always the same as the
vertIndex, so the data structure holds redundant information. (Because of this, one could use a more
streamlined data structure for such meshes. What would it be?) Polygonalization of the sphere in this
way is straightforward, but for more complicated shapes it can be very tricky. See [refs] for further discussions.
Ultimately we need a method, such as makeSurfaceMesh(), that generates such meshes for a given
surface P(u, v). We discuss the implementation of such a function in Case Study 6.13.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 37

Note that some graphics packages have routines that are highly optimized when they operate on triangles. To exploit these we might choose to polygonalize the sphere into a collection of triangles, subdividing each quadrilateral into two triangles.

A simple approach would use the same vertices as above, but alter the face list replacing each quadrilateral with two triangles. For instance, a face that uses vertices 2, 3, 15, 14 might be subdivided into two
triangles, once using 2, 3, 15 and the other using 2, 15, 14.
The sphere is a special case of a surface of revolution, which we treat in Section 6.5.7. The tapered cylinder is also a surface of revolution. It is straightforward to develop a mesh model for the tapered cylinder.Figure 6.59 shows the tapered cylinder approximated with nSlices = 10 and nStacks = 1. A
decagon is used for its cap and base. (If you prefer to use only triangles in a mesh, the walls, the cap, and
the base could be dissected into triangles. (How?))

Figure 6.59. A mesh approximation to the tapered cylinder.

Practice Exercises.
6.5.8. The mesh for a given sphere. Write out the vertex, normal, and face lists for the sphere, when
nSlices = 6 and nStacks = 4, choosing a convenient numbering scheme.
6.5.9. Restricting the mesh to triangular faces. Adjust the lists in the previous exercise for the case
where all faces are triangles.
6.5.10. The mesh for a cylinder and cone. Write out vertex, normal, and face lists for the generic tapered cylinder that uses nSlices = 4 and nStacks = 2.
6.5.6. Ruled Surfaces.
We resume an exploration of curved surfaces with the family of ruled surfaces. This family is simple to
describe yet provides a wide variety of useful and interesting shapes. We study how to describe them,
polygonalize them and how to compute the normal vector to them at each point.
Ruled surfaces (also called "lofted surfaces") are swept out by moving a straight line along a particular
trajectory. They are composed of a collection of straight lines in the following sense:
Definition: A surface is ruled if through every one of its points there passes at least one line that lies
entirely on the surface.
Because ruled surfaces are based on a family of lines, it is not surprising to find buried in their parametric representations something akin to the familiar form for a line, P(v) = (1 - v) P0+ v P1, where P0 and P1

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 38

are points. But for ruled surfaces the points P0 and P1 become functions of another parameter u: P0 becomes P0(u), and P1 becomes P1(u). Thus the ruled surfaces that we examine have the parametric form
P(u, v) = (1 - v) P0(u) + v P1(u)

(6.35)

The functions P0(u) and P1(u) define curves lying in 3D space. Each is described by three component
functions, as in P0(u) = (X0(u), Y0(u), Z0(u)). Both P0(u) and P1(u) are defined on the same interval in u
(commonly from 0 to 1). The ruled surface consists of one straight line joining each pair of corresponding points, P0(u) and P1(u), for each u in (0,1), as indicated in Figure 6.60. At v = 0 the surface is at
P0(u), and at v = 1 it is at P1(u). The straight line at u = u is often called the ruling at u.
1st Ed. Figure 9.3
Figure 6.60. A ruled surface as a family of straight lines.
For a particular fixed value, v the v-contour is some blend of the two curves P0(u) and P1(u). It is an affine combination of them, with the first weighted by (1 - v) and the second by v. When v is close to 0,
the shape of the v-contour is determined mainly by P0(u), whereas when v is close to 1, the curve P1(u)
has the most influence.
If we restrict v to lie between 0 and 1, only the line segment between corresponding points on the curves
will be part of the surface. On the other hand, if v is not restricted, each line will continue forever in both
directions, and the surface will resemble an unbounded curved sheet. A ruled patch is formed by restricting the range of both u and v, to values between, say, 0 and 1.
A ruled surface is easily polygonalized in the usual fashion: choose a set of samples ui and vj, and compute the position P(ui, vj) and normal n(ui, vj) at each. Then build the lists as we have done before.
Some special cases of ruled surfaces will reveal their nature as well as their versatility. We discuss three
important families of ruled surfaces, the cone, the cylinder, and the bilinear patch.
Cones.
A cone is a ruled surface for which one of the curves, say, P0(u), is a single point P0(u) = P0, the apex
of the cone, as suggested in Figure 6.61a. In Equation 6.35 this restriction produces:

Figure 6.61. A cone.

P(u, v) = (1 - v) P0 + v P1(u)

{a general cone}

(6.36)

For this parameterization all lines pass through P0 at v = 0, and through P1(u) at v = 1. Certain special
cases are familiar. A circular cone results when P1(u) is a circle, and a right circular cone results when
the circle lies in a plane that is perpendicular to the line joining the circle's center to P0. The specific example shown in Figure 6.61b uses P1(u) = (r(u) cos u, r(u) sin u, 1) where the radius curve r(u) varies
sinusoidally: r(u) = 0.5 + 0.2 cos(5u).
Cylinders.
A cylinder is a ruled surface for which P1(u) is simply a translated version of P0(u): P1(u) = P0(u) + d,
for some vector d, as shown in Figure 6.62a. Sometimes one speaks of sweeping the line with endpoints P0(0) and P0(0) + d (often called the generator) along the curve P0(u) (often called the directrix), without altering the direction of the line.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 39

Figure 6.62*. a). A cylinder. b). ribbon candy cylinder.

The general cylinder therefore has the parametric form
P(u, v) = P0(u) + dv

(6.37)

To be a true cylinder, the curve P0(u) is confined to lie in a plane. If P0(u) is a circle the cylinder is a circular cylinder. The direction d need not be perpendicular to this plane, but if it is, the surface is called a
right cylinder. This is the case for the generic cylinder. Figure 6.55b shows a ribbon candy cylinder
where P0(u) undulates back and forth like a piece of ribbon. The ribbon shape is explored in the exercises.
Bilinear Patches.
A bilinear patch is formed when both P0(u) and P1(u) are straight line segments defined over the same
interval in u, say, 0 to 1. Suppose the endpoints of P0(u) are P00 and P01, (so that P0(u) is given by (1 u)P00 + u P01), and the endpoints of P1(u) are P10 and P11. Then using Equation 6.35 the patch is given parametrically by:
P(u, v) = (1 - v)(1 - u)P00 + (1 - v) u P01 + v(1 - u)P10 + uv P11

(6.38)

This surface is called bilinear because its dependence is linear in u and linear in v. Bilinear patches
need not be planar; in fact they are planar only if the lines P0(u) and P1(u) lie in the same plane (see the
exercises). Otherwise there must be a twist in the surface as we move from one of the defining lines to
the other.
An example of a nonplanar bilinear patch is shown in Figure 6.63. P0(u) is the line from (2, -2, 2) to (2,
2, -2), and P1(u) is the line from (-2,-2,-2) to (-2, 2, 2). These lines are not coplanar. Several u-contours
are shown, and the twist in the patch is clearly visible.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 40

Figure 6.63. A bilinear patch.

The normal vector to a bilinear patch is easily found using Equation 6.26. If the patch is planar the direction of the normal is constant but its magnitude can vary with u and v. If the patch is nonplanar both
the magnitude and direction of the normal vector vary with position.
Other ruled surfaces.
There are many other interesting ruled surfaces. Figure 6.64a shows a double helix formed when P0(u) and P1(u)
are both helixes that wind around each other. Part b shows the intriguing Mbius strip that has only one edge. The
exercises explore the parametric representation of these surfaces. Part c shows a vaulted roof made up of four
ruled surfaces. Case Study 6.8 examines modeling such vaulted domes of a cathedral.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 41

Figure 6.64. a). double helix b). Mbius strip, c).Vaulted roof.
Bilinearly blended Surfaces - Coons Patches.
An interesting and useful generalization of a ruled surface, which interpolates two boundary curves P0(u)
and P1(u), is a bilinearly blended patch that interpolates to four boundary curves. This family was first
developed by Steven Coons [coons] and is sometimes called a Coons patch.
Figure 6.65 shows four adjoining boundary curves, named pu0(u), pu1(u), p0v(v), and p1v(v). These curves
meet at the patch corners (where u and v are combinations of 0 and 1) but otherwise have arbitrary
shapes. This therefore generalizes the bilinear patch for which the boundary curves are straight lines.

Figure 6.65. Four boundary curves determining a Coons patch.

We want a formula P(u, v) that produces a smooth transition from each boundary curve to the other as u
and v vary.
A natural first guess is to somehow combine a ruled patch built out of pu0(u) and pu1(u), with a ruled
patch built out of p0v(v) and p1v(v). But simply adding such surfaces doesn't work: it fails to interpolate
the four curves properly (and it's illegal: it results in a non-affine combination of points!). The trick is to
add these surfaces and then subtract the bilinear patch formed from the four corners of these curves. Figure 6.66 shows visually how this works [heckbert94].
surf 1 + surf 2 - surf 3 = coons patch
Figure 6.66. Combining patches to create the Coons patch.
The formula for the patch is therefore

P(u, v ) = [ p0 v (v) (1 - u) + p1v (v ) u] + [ pu0 (u)(1 - v) + pu1 (u) v]

-[(1 - u)(1 - v ) p0 v (0 ) + u(1 - v) p1v (0) u + p0 v (1)v (1 - u) + p1v (1)u v]

(6.39)

Note that at each (u, v) this is still an affine combination of points, as we insist. Check that at (u, v)=(0,
0) this evaluates to pu0(0), and similarly that it coincides with the other three corners at the other extreme
values of u and v. Figure 6.67 shows an example Coons patch bounded by curves that have a sinusoidal
oscillation.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 42

Figure 6.67. Example of a Coons patch.

Practice Exercises.
6.5.11. A Pyramid is a cone. What shape should P1(u) have to create a ruled surface that is a pyramid
with a square base? Give specific expressions for the curve and the point P0 so that the square base of the
pyramid lies in the x, z-plane, centered at the origin, with sides of length 2. The pyramid should have
height 1.5.
6.5.12. Ribbon candy cylinders. Find a parametric form for P0(u) that produces a good approximation to
the ribbon candy cylinder shown in Figure 6.62b. Consider the ribbon as wrapped about a succession of
abutting circular cylinders of radius 1. The center of the i-th cylinder lies at (xi, yi) = (id, r), where the
+ and - alternate, and d2 = 1 - r2. Choose some value of r between 0 and 1.
6.5.13. The double helix. The parametric form for a helix is (cos(t), sin(t), t). Find expressions for two
helices, P0(u) and P1(u), both of which wind around the z-axis yet are 180o out of phase so that they wind
around each other. Write the parametric form for the ruled surface formed using these two curves.
6.5.14. The Mbius strip. Find a parametric form for the Mbius strip shown in Figure 6.64b.
Hint: revolve a line about the z-axis, but put in a twist as it goes around. Does the following attempt do
the job: Po(u) = (cos(2pu)), sin(2pu), u), and P1(u) = (cos(2pu)), sin(2pu), 1-u)?
6.5.15. Is it affine? Show that the Coons patch P(u,v) of Equation 6.33 is composed of an affine combination of points.
6.5.16. Does it really interpolate? Check that P(u, v) of Equation 6.33 interpolates each of the four
boundary curves, and therefore interpolates each of the four corners.
6.5.7. Surfaces of Revolution.
As described earlier, a surface of revolution is formed by a rotational sweep of a profile curve, C,
around an axis. Suppose we place the profile in the xz-plane and represent it parametrically by C(v) =
(X(v), Z(v)). To generate the surface of revolution we sweep the profile about the z-axis under control of
the u parameter, with u specifying the angle through which each point has been swept about the axis. As
before, the different positions of the curve C around the axis are called meridians. When the point (X(v),
0, Z(v)) is rotated by u radians, it becomes (X(v)cos(u), X(v)sin(u), Z(v)). Sweeping it completely around
generates a full circle, so contours of constant v are circles, called parallels of the surface7. The parallel
at v has radius X(v) and lies at height Z(v) above the xy-plane. Thus the general point on the surface is
7

More formally, a meridian is the intersection of the surface with a plane that contains the axis of revolution.,
and a parallel is the intersection of the surface with a plane perpendicular to the axis.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 43

P(u, v) = ( X (v ) cos(u), X (v ) sin(u), Z (v ))

(6.40)

The generic sphere, tapered cylinder, and cone are all familiar special cases. (What are their profiles?)
The normal vector to a surface of revolution is easily found by direct application of Equation 6.34 to
Equation 6.19 (see the exercises). This yields

n(u, v ) = X (v ) Z (v ) cos(u), Z (v ) sin(u), X (v )

(6.41)

where the dot denotes the first derivative of the function. The scaling factor X(v) disappears upon normalization of the vector. This result specializes to the forms we found above for the simple generic
shapes (see the exercises).
For example, the torus is generated by sweeping a displaced circle about the z-axis, a shown in Figure
6.68. The circle has radius A and is displaced along the x-axis by D, so that its profile is C(v) = (D + A
cos(v), A sin(v)). Therefore the torus (Figure 6.68b) has representation

Figure 6.68. A torus.

P(u, v) = ((D + A cos(v)) cos(u), (D + A cos(v)) sin(u), A sin(v))

(6.42)

Its normal vector is developed in the exercises.

We usually sweep a curve lying in a plane about an axis that lies in that plane, but a surface of revolution
can be formed by sweeping about any axis. Choosing different axes for a given profile can lead to interesting families of surfaces. The general form for P(u, v) for the surface of revolution about an arbitrary
axis is developed in the exercises.
A mesh for a surface of revolution is built in a program in the usual way (see Section 6.5.4). We choose
a set of u and v values, {ui} and {vj}, and compute a vertex at each from P(ui, vj), and a normal direction
from n(ui, vj). Polygonal faces are built by joining four adjacent vertices with straight lines. A method to
do this is discussed in Case Study 6.13.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 44

Figure 6.69 shows another example in which we try to model the dome of the exquisite Taj Mahal in
Agra, India, shown in part a. Part b shows the profile curve in the xz-plane, and part c shows the resulting
surface of revolution. Here we describe the profile by a collection of data points Ci = (Xi, Zi), since no
suitable parametric formula is available. (We rectify this lack in Chapter 8 by using a B-spline curve to
form a smooth parametric curve based on a set of data points.)

a).

c).

b).

Figure 6.69*. A surface of revolution - the dome of the Taj Mahal.

To build a surface of revolution when the profile consists of discrete points, simply take as the ij-th vertex the slightly different form of Equation 6.41:
Pi,j = (Xj cos(ui), Xj sin(ui), Zj)
Practice Exercises.
6.5.17. The generic shapes as surfaces of revolution. Describe the profiles of the generic sphere, cylinder, and cone parametrically, and then express them as surfaces of revolution.
6.5.18. Rotation About Other Axes. Consider a profile curve C(v) = (X(v), Z(v)), lying in the xz-plane,
and an arbitrary axis through the origin given by unit vector r. We know from Equation 5.33 that the
matrix Rr() performs a rotation of a point through radians about the axis r.
a). From this show that the surface of revolution formed by sweeping C(v) about axis r is

X(v)
0
( X (u, v), Y (u, v ), Z (u, v),1) = R (u)
Z(v)
1
r

b). Check this for the special case of rotation about the z-axis.
c). Repeat part b for rotations about the x-axis, and about the y-axis.
6.5.19. Finding normal vectors. a). Apply Equation 6.40 to Equation 6.25 to derive the form in Equation 6.41 for the normal vector to a surface of revolution. b). Use this result to find the normal to each of

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 45

the generic sphere, cylinder, and cone, and show the results agree with those found in Section 6.5.3.
Show that the normal vector to the torus has the form
n(u, v) = (cos(v)cos(u), cos(v)sin(u), sin(v))(D + A cos(v)). Also, find the inside-outside function for the
torus, and compute the normal using its gradient.
6.5.20. An Elliptical Torus. Find the parametric representation for the following two surfaces of revolution: a). The ellipse given by (a cos(v), b sin(v)) is first displaced R units along the x-axis and then revolved about the y-axis. b). The same ellipse is revolved about the x-axis.
6.5.21. A Lissajous of Revolution. Sketch what the surface would look like if the Lissajous figure of
Equation 6.19 with M = 2, N = 3, and = 0 is rotated about the y-axis.
6.5.22. Revolved n-gons. Sketch the surface generated when a square having vertices (1, 0, 0), (0, 1, 0),
( - 1, 0, 0), (0, - 1, 0) is revolved about the y-axis. Repeat for a pentagon and a hexagon.
6.5.8. The Quadric Surfaces.
An important family of surfaces, the quadric surfaces, are the 3D analogs of conic sections (the ellipse,
parabola, and hyperbola, which we examined in Chapter 3. Some of the quadric surfaces have beautiful
shapes and can be put to good use in graphics.
The six quadric surfaces are illustrated in Figure 6.70.
1st Ed. Figure 9.12. the 6 quadric surfaces a),f).
Figure 6.70. The six quadric surfaces: a. Ellipsoid, b. Hyperboloid of one sheet, c. Hyperboloid of two
sheets, d. Elliptic cone, e. Elliptic paraboloid, f. Hyperbolic paraboloid.
We need only characterize the generic versions of these shapes, since we can obtain all the variations
of interest by scaling, rotating, and translating the generic shapes. For example, the ellipsoid is usually
said to have the insideoutside function
2
2
2
x
y
z
(6.43)
F ( x, y, z) = + + 1
a
b
c
so that it extends in x from -a to a, in y from -b to b, and in z from -c to c. This shape may be obtained
from the form of the generic sphere by scaling it in x, y, and z by a, b, and c, respectively, as we describe
below. We can obtain rotated versions of the ellipsoid in a similar manner.
Figure 6.71 provides descriptions of the six generic quadric surfaces, giving both their implicit and
parametric forms. We discuss some interesting properties of each shape later.
name of
Insideparametric form
v-range,u-range
quadric
outside function
ellipsoid
( / 2, / 2),( ,
x 2 + y 2 + z 2 1 (cos(v ) cos( u),cos( v ) sin( u),

sin( v ))
hyperboloid
of one sheet

x 2 + y2 z2 1

(sec(v ) cos( u), sec( v ) sin( u),

tan( v ))

( / 2, / 2),( ,

hyperboloid
of two
sheets

x2 y2 z2 1

(sec(v ) cos(u), sec(v ) tan( u),

tan( v ))

( / 2, / 2) 8

elliptic cone

x 2 + y 2 z2

(v cos( u), v sin(u), v )

any real, ( , )

The v-range for sheet #1 is (-/2,/2), and for sheet #2 is (/2,3/2).

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 46

elliptic
paraboloid

x2 + y2 z

(v cos( u), v sin(u), v 2 )

v 0,( , )

hyperbolic
paraboloid

x 2 + y2 z

(v tan(u), v sec(u), v 2 )

v 0,( , )

Figure 6.71. Characterization of the six generic quadric surfaces.

Figure 6.64 also reports the parametric form for each of the generic quadric surfaces. It is straightforward
to check that each of these forms is consistent with its corresponding implicit form: substitute the parametric form for the x-, y-, and z- components, and use trigonometric identities to obtain 0.
Note that a change of sign in one term of the insideoutside function causes a cos( ) to be changed to
sec( ) and a sin( ) to be changed to tan( ) in the parametric forms. Both sec( ) and tan( ) grow without
bound as their arguments approach /2, and so when drawing u-contours or v-contours the relevant parameter is restricted to a smaller range.
Some Notes on the Quadric Surfaces.
We shall summarize briefly some significant properties of each quadric surface. One such property of a
quadric is the nature of its traces. A trace is the curve formed when the surface is cut by a plane. All
traces of a quadric surface are conic sections - see the exercises. The principal traces are the curves
generated when the cutting planes are aligned with the axes: the planes z = k, y = k, or x = k, where k is
some constant.
In the discussion that follows we suppose that the generic surfaces have been scaled in x-, y-, and z- by
values a, b, and c, respectively, to make it easier to talk about the dimensions of the surfaces, and to distinguish cases where the surface is a surface of revolution or not.
Ellipsoid. Compare the implicit form and parametric form for the ellipse in Chapter 3 to see how they
extend the 2D ellipse to this 3D ellipsoid. Parameters a, b, and c give the extent of the ellipsoid along
each axis. When two of the parameters are equal, the ellipsoid is a surface of revolution. (If a = b, what
is the axis of revolution?). When a, b, and c all are equal, the ellipsoid becomes a sphere. All traces of
the ellipsoid are ellipses.
Hyperboloid of one sheet. When a = b, the hyperboloid becomes a surface of revolution formed by
rotating a hyperbola about an axis. The principal traces for the planes z = k are ellipses, and those for the
planes x = k and y = k are hyperbolas. The hyperboloid of one sheet is particularly interesting because it
is a ruled surface, as suggested in Figure 6.72a. If a thread is woven between two parallel ellipses, as
shown, this surface will be created. Formulas for the rulings are discussed in the exercises.
9.13 a). hyperboloid of one sheet
b). hyperbolic paraboloid.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 47

Figure 6.72. Two ruled quadric surfaces.

Hyperboloid of two sheets. No part of the surface lies between x = -a and x = a. (Why?). When a = b,
it becomes a surface of revolution. The traces for planes x = k when |k| > a are ellipses, and the other
principal traces are hyperbolas.
Elliptic cone. The elliptic cone is a special case of the general cone treated earlier: Its generator lines
trace an ellipse. This cone is, of course, a ruled surface, and the principal traces for planes z =k are ellipses. What are traces for planes that contain the z-axis? When a = b, this quadric is a surface of revolution: It becomes a right circular cone.
Elliptic paraboloid. The traces of an elliptic paraboloid for planes z = k > 0 are ellipses, and the other
principal traces are parabolas. When a = b, it is a surface of revolution.
Hyperbolic paraboloid. The hyperbolic paraboloid is sometimes called a saddle-shaped surface.
The traces for planes z = k (when k 0) are hyperbolas, and for planes x = k or y = k they are parabolas.
(What is the intersection of this surface with the plane z = 0?) This is also a ruled surface (see Figure
6.65b).
Normal Vectors to Quadric Surfaces.
Because the implicit form for each quadric surface is quadratic in x, y, and z, taking the gradient to find
the normal presents no problem. Further, since each component of the gradient vector is linear in its own
variable or just a constant, it is straightforward to write the gradient in parametric form: just substitute
X(u, v) for x, etc.
For instance, the gradient of F(x, y, z) for the ellipsoid is

F = (2 x ,2 y ,2 z )
and so the normal in parametric form is (after deleting the 2)

n(u, v ) = (cos(v ) cos(u),cos( v ) sin( u),sin(v ))

(6.44)

Normals for the other quadric surfaces follow just as easily, and so they need not be tabulated.
Practice exercises.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 48

6.5.23. The Hyperboloid is a Ruled Surface. Show that the implicit form of a hyperboloid of one sheet
can be written (x + z)(x - z) =(1 - y)(1 + y). Show that therefore two families of straight lines lie in the
surface: the family x - z = A(1 - y) and the family A(x + z) = 1 + y, where A is a constant. Sketch these
families for various values of A. Examine similar rulings in the hyperbolic paraboloid.
6.5.24. The hyperboloid of one sheet. Show that an alternative parametric form for the hyperboloid of
one sheet is p(u,v) =(cosh(v) cos(u), cosh(v) sin(u, sinh(v)).
6.5.25. Traces of Quadrics are Conics. Consider any three (noncollinear) points lying on a quadric surface. They determine a plane that cuts through the quadric, forming the trace curve. Show that this curve
is always a parabola, ellipse, or hyperbola.
6.5.26. Finding Normals to the Quadrics. Find the normal vector in parametric form for each of the six
quadric surfaces.
6.5.27. The Hyperboloid As a Ruled Surface. Suppose (x0, y0, 0) is a point on the hyperboloid of one
sheet. Show that the vector R ( t ) = ( x0 + y0 t , y0 x0 t , t )
describes a straight line that lies everywhere on the hyperboloid and passes through (x0, y0, 0). Is this sufficient to make the surface a ruled surface? Why or why not? [apostol p.329]
6.5.28. The Hyperbolic Paraboloid As a Ruled Surface. Show that the intersection of any plane parallel to the line y = x cuts the hyperbolic paraboloid along a straight line.
6.5.9. The Superquadrics.
Following the work of Alan Barr [barr81], we can extend the quadric surfaces to a vastly larger family,
in much the way we extended the ellipse to the superellipse in Chapter 3. This provides additional interesting surface shapes we can use as models in applications.
Barr defines four superquadric solids: the superellipsoid, superhyperboloid of one sheet, and superhyperboloid of two sheets, which together extend the first three quadric surfaces, and the supertoroid which
extends the torus. The extensions introduce two bulge factors, m and n, to which various terms are
raised. These bulge factors affect the surfaces much as n does for the superellipse. When both factors
equal 2, the first three superquadrics revert to the quadric surfaces catalogued previously. Example
shapes for the four superquadrics are shown in Figure 6.73.
1st Ed. Figure 9.15
Figure 6.73. Examples of the four superquadrics. n1 and n2 are (left to right): 10, 2, 1.11, .77, and .514.
(Courtesy of Jay Greco.)
The inside-outside functions and parametric forms9 are given in Figure 6.74.
name of quadric
Implicit form
parametric form
n
n m/ n
m
(cos2 / m ( v ) cos2 / n ( u), cos2 / m ( v ) sin 2 / n ( u),
superellipsoid
x + y % + z 1

v-range,u-range
[ / 2, / 2],[ , ]

sin 2 / m ( v ))

superhyperboloid
of one sheet

xn + yn %

m/ n

superhyperboloid
of two sheets

xn yn %

m/n

supertoroid

1/ n

xn yn %

zm 1

(sec2 / m ( v ) cos 2 / n ( u), sec 2 / m ( v ) sin 2 / n ( u),

zm 1

(sec 2 / m (v ) sec 2 / n ( u), sec2 / m ( v ) tan 2 / n ( u),

D' + z m 1

( / 2, / 2),[ , ]

tan 2 / m ( v ))

( / 2, / 2 )

tan 2 / m ( v ))
(( d + cos2 / m ( v )) cos2 / n ( u ),( d + cos2 / m ( v )) sin 2 / n ( u),

[ , ),[ , )

sin 2 / m ( v))

Figure 6.74. Characterization of the four superquadric surfaces.

Keep in mind that it's illegal to raise a negative value to a fractional exponent. So expressions such as
2/m
2/m-1
cos (v) should be evaluated as cos(v)|cos(v)| or the equivalent.
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 49

These are generic superquadrics in the sense that they are centered at the origin, aligned with the coordinate axes, and have unit dimensions. Like other shapes they may be scaled, rotated, and translated as
desired using the current transformation to prepare them properly for use in a scene.
The normal vector n(u, v) can be computed for each superquadric in the usual ways. We shall give only
the results here.
Normals to the superellipsoid and supertoroid.
The normal vectors for the superellipsoid and the supertoroid are the same:

n(u, v ) = (cos2 2 / m (v ) cos2 2 / n (u), cos2 2/ m (v ) sin 2 2 / n (u), sin 2 2 / m (v ))

(6.45)

(How can they be the same? The two surfaces surely dont have the same shape.)
The Superhyperboloid of one sheet.
The normal vector to the superhyperboloid of one sheet is the same as that of the superellipsoid, except
all occurrences of cos(v) are replaced with sec(v) and those of sin(v) are replaced with tan(v). Do not alter cos(u) or sin(u) or any other term.
The Superhyperboloid of two sheets.
For the superhyperboloid of two sheets, the trigonometric functions in both u and v are replaced. Replace
all occurrences of cos(v) with sec(v), those of sin(v) with tan(v), those of cos(u) with sec(u), and those of
sin(u) with tan(u).
Practice Exercises.
6.5.29. Extents of Superquadrics. What are the maximum x, y, and z values attainable for the superellipsoid and the supertoroid?
6.5.30. Surfaces of revolution. Determine the values of bulge factors m and n for which each of the superquadrics is a surface of revolution, and find the axis of revolution. Describe any other symmetries in
the surfaces.
6.5.31. Deriving the Normal Vectors. Derive the formula for the normal vector to each superquadric.
surface.
6.5.10. Tubes Based on 3D Curves.
In Section 6.4.4 we studied tubes that were based on a spine curve C(t) meandering through 3D space.
A polygon was stationed at each of a selection of spine points, and oriented according to the Frenet
frame computed there. Then corresponding points on adjacent polygons were connected to form a flatfaced tube along the spine.
Here we do the same thing, except we compute the normal to the surface at each vertex, so that smooth
shading can be performed. Figure 6.75 shows the example of a tube wrapped around a helical shape.
Compare this with Figure 6.47.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 50

Figure 6.75. A helical tube undulating through space.

If we wish to wrap a circle (cos(u), sin(u), 0) about the spine C(t), the resulting surface has parametric
representation

P(u, v ) = C(v) + cos(u)N(v ) + sin(u)B(v)

(6.46)

where the normal vector N(t) and binormal vector B(t) are those given in Equations 6.14 and 6.15. Now
we can build a mesh for this tube in the usual way, by taking samples of P(u, v), building the vertex,
normal, and face lists, etc. (What would be altered if we wrapped a cycloid - recall Figure 3.77 - about
the spine instead of a circle?)
6.5.11. Surfaces based on Explicit Functions of Two Variables.
Many surface shapes are single-valued in one dimension, so their position can be represented as an explicit function of two of the independent variables. For instance, there may be a single value of height
of the surface above the xz-plane for each point (x, z), as suggested in Figure 6.76. We can then say that
the height of the surface at (x, z) is some f(x, z). Such a function is sometimes called a height field
[bloomenthal97]. A height field is often given by a formula such as

Figure 6.76. A single-valued height field above the xz-plane.

f ( x, z ) = e ax

by 2

(6.47)

(where a and b are given constants), or the circularly symmetric sinc function
. f ( x, z ) =

sin( x 2 + z 2 )

(6.48)

x2 + z2

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 51

Contrast this with surfaces such as the sphere, for which more than one value of y is associated with each
point (x, z). Single valued functions permit a simple parametric form:
P(u, v) = (u, f(u, v), v)

(6.49)

and their normal vector is n(u, v) = ( f / u,1, f / v) (check this). That is, u and v can be used
directly as the dependent variables for the function. Thus u-contours lie in planes of constant x, and vcontours lie in planes of constant z. Figure 6.77a shows a view of the example in Equation 6.47, and Figure 6.77b shows the function of Equation 6.48.

Figure 6.77. Two height fields. a) gaussian, b). sinc function. (files: fig6.77a.bmp, fig6.77b.bmp)
Each line is a trace of the surface cut by a plane, x = k or z = k, for some value of k. Plots such as these
can help illustrate the behavior of a mathematical function.
Practice exercise 6.5.32. The Quadrics as Explicit Functions. The elliptic paraboloid can be written as
z = f(x, y), so it has an alternate parametric form (u, v, f(u, v)). What is f()? In what ways is this alternate
parametric form useful? What other quadrics can be represented this way? Can any superquadrics be represented explicitly this way?

6.6. Summary
Hill - ECE660:Computer Graphics Chapter 6

w v , and that v w = 2 Area(T )m .

We now explore the projection of T onto the plane with normal vector n.

n
v

w'
T'

v'
Figure 6.80. Effect of Orthographic projection on area.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 55

This projection T is defined by the projected vectors w and v: its area Area( T' ) =

v' w' = 2 Area(T' )n . We now calculate w and v, and form Area(T).

1
2

w' v' , so

b). Using ideas from Chapter 4, show that that v projects to v given by v = v - (v n) n and similarly that w
= w - (w n) n.
So we need only relate the sizes of the two cross products.
c). Use the forms of u and v to show that: v w = v w - (w n) (v n) + (v n)(w n) + (w n)(v
n)(n n)
and explain why the last term vanishes. Thus we have
2 Area(T) n = v w - (w n) (v n) + (v n)(w n)
d). Dot both sides of this with n and show that the last two terms drop out, and that we have 2 Area(T) = v
w n = 2Area(T) m n as claimed.
e). Show that this result generalizes to the areas of any planar polygon P and its projected image P.
f). Recalling that a dot product is proportional to the cosine of an angle, show that Area(T) = Area(T) cos
and state what the angle is.
g). Show that the areas Ax, Ay, Az defined above are simply Kmx, Kmy, and Kmz, respectively, where K is some
constant. Hence the areas Ax, Ay, and Az are in the same ratios as mx, my, and mz.
B). So to find m we need only compute the vector (Ax, Ay, Az), and normalize it to unit length. We now show
how to compute the area of the projection of P of Equation 6.50 onto the xy-plane directly from its vertices.
The other two projected areas follow similarly.
Each 3D vertex Pi = (xi, yi, zi ) projects onto the xy-plane as Vi = (xi, yi). Figure 6.81 shows an example projected polygon P.

Figure 6.81. Computing the area of a polygon.

Each edge of P defines a trapezoidal region lying between it and the x-axis.
h). Show that the area of such a trapezoid is the width of the base times the midpoint of the edge. For instance
the area A0 in the figure is A0 = 0.5(x0 - x1)(y0 + y1). This quantity is negative if x0 lies to the left of x1, and is
positive otherwise.
We use this same form for each edge. For the i-th edge define

Ai =

1
( x i x next (i ) )( y i + y next (i ) )
2

where next(i) is 0 if i is equal to N-1, and is i+1 otherwise.

i). Show that if two adjacent edges of the polygon are collinear (which would make a cross product based on
them zero), the areas contributed by these edges is still properly accounted for.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 56

j). Show that the sum of the Ai properly adds the positive and negative contributions of area to form a resultant
sum that is either the area of the polygon, or its negative.
Now we ask which of the two basic directions does m point in? That is, if you circle the fingers of your right
hand around the vertices of the polygon moving from P0 to P1, to P2 etc., the direction of the arrow in Figure
6.82, does m point along your thumb or in the opposite direction?

z
P1
m

x
Figure 6.82. The direction of the normal found by the Newell method.
j). Show that m does point as shown in Figure 6.82. Thus for a mesh that has a well-defined inside-outside, we
can say that m is the outward pointing normal if the vertices are labeled CCW as seen from the outside.
6.7.3. Case Study 6.3. The Prism.
(Level of Effort: III) Write an application that allows the user to specify the polygonal base of a prism using
the mouse. It then creates the vertex, normal, and face lists for the prism, and displays it.
Figure 6.83a shows the user's drawing area, a square presented on the screen. The user lays down a sequence
of points in this square with the mouse, terminating the process with a right click.

a).

b).
z

1
1
x

Figure 6.83. Designing and building a prism mesh.

In 3D space the corresponding square is considered to be a unit square lying in the xy-plane as suggested in
Figure 6.83b, and the base of the prism lies within it. This establishes the size of the base in the 3D world.
The prism is considered to be the base polygon after it is swept one unit in the direction of the z-axis. Exercise
the program on several prisms input by the user. See if your implementation of OpenGL properly draws nonconvex base polygons.
6.7.4. Case Study 6.4. Prism Arrays and Extruded Quad-strips.
(Level of Effort: III) Write the two methods described in Section 6.4.2:
void Mesh:: makePrismArray(<... suitable arguments ..>);
void Mesh:: makeExtrudedQuadStrip(Point2 p[], int numPts, Vector3 d);
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 57

that create meshes for an array of prisms, and for an extruded quad-strip.
a). Arrays of prisms: Choose an appropriate data type to represent an array of prisms. Note that makePrismArray() is similar to the method that makes a mesh for a single prism. Exercise the first method on at least
the two block letters with the shapes K and W. (Try D also if you wish.)
b). Extruded quad-strips used to form tubes: The process of building the vertex, normal, and face lists of a
mesh is really a matter of keeping straight the many indices for these arrays. To assist in developing this
method, consider a quad-strip base polygon described as in Equation 6.9 by the vertices
quad-strip = {p0, p1, ..., pM-1}
where pi = (xi, yi, 0) lies in the xy-plane, as shown in Figure 6.84a. When extruded, each successive pair of
vertices forms a waist of the tube, as shown in Figure 6.84b. There are num = M/2 - 1 segments in the tube.
a). quad-strip in xy-plane b) the four extruded segments
Figure 6.84. Building a mesh from a quad-strip base polygon.
The 0-th waist consists of vertices p0, p1, p1 + d, and p0 + d, where d is the extrusion vector. We add vertices to
the vertex list as follows: pt[4i] = p2i, pt[4i + 1] = p2i+1, pt[4i + 2] = p2i+1 + d, and pt[4i+3] = p2i + d,
for i = 0, ..., num as suggested in Figure 6.84b.
Now for the face list. We first add all of the outside walls of each segment of the tube, and then append the
end walls (i.e. the first end wall uses vertices of the first waist). Each of the num segments has four walls.
For each wall we list the four vertices in CCW order as seen from the outside. There are patterns in the various
indices encountered, but they are complicated. Check that the following vertex indices are correct for each of
the four walls of the k-th segment: The j-th wall of the k-th segment has vertices with indices: i0, i1, i2, and i3,
where:
i0 = 4k+j
i1 = i0 + 4
i3 = 4k + (j + 3) % 4
i2 = i3 + 4;
for k = 0, 1, .., num and j = 0, 1, 2, 3.
What are indices of the two end faces of the tube?
Each face has a normal vector determined by the Newell method, which is straightforward to calculate at the
same time the vertex indices are placed in the face list. All vertex normals of a face use the same normal vector: face[L].normindex = {L,L,L,L}, for each L.
Exercise the makeExtrudedQuadStrip() method by modeling and drawing some arches, such as the one
shown in Figure 6.39, as well as some block letters that permit the use of quad-strips for their base polygon.
6.7.5. Case Study 6.5. Tubes and Snakes based on a Parametric Curve.
(Level of Effort III) Write and test a method
void Mesh:: makeTube(Point2 P[], int numPts, float t[], int numTimes)
that builds a flat-faced mesh based on wrapping the polygon with vertices P0, P1, ..., PN-1 about the spine curve
C(t). The waists of the tube are formed on the spine at the set of instants t0, t1, ..., tM-1 , and a Frenet frame is
constructed at each C(ti). The function C(t) is hard-wired into the method as a formula, and its derivatives
are formed numerically.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 58

Experiment with the method by wrapping polygons taken from Example 3.6.3 that involve a line jumping back
and forth between two concentric circles. Try at least the helix and a Lissajous figure as example spine curves.
6.7.6. Case Study 6.6. Building Discrete-Stepped Surfaces of Revolution.
(Level of Effort: III) Write an application that allows the user to specify the profile of an object with the
mouse, as in Figure 6.85. It then creates the mesh for the surface of revolution, and displays it. The program
also writes the mesh data to a file in the format described in Case Study 6.1.
Figure 6.85a shows the user's drawing area, a square presented on the screen. The user lays down a sequence
of points in this square with the mouse.

Figure 6.85. Designing a profile for a surface of revolution.

In 3D space the corresponding square is considered to be a unit square lying in the xy-plane. This establishes
the size of the profile in the 3D world. The surface of revolution is formed by sweeping the profile about the
z-axis, in a user-defined number of steps.
Exercise the program on several surfaces of revolution input by the user.
6.7.7. Case Study 6.7. On Edge Lists and Wireframe Models.
(Level of Effort: II) A wireframe version of a mesh can be drawn by drawing a line for each edge of the mesh.
Write a routine void Mesh:: drawEdges(void) that does this for any given mesh. It simply traverses
each face, connecting adjacent vertices with a line. This draws each line twice. (Why?)
In some time critical situations this inefficiency might be intolerable. In such a case an edge list can be built
for the mesh which contains each edge of the mesh only once. An edge list is an array of index pairs, where
the two indices indicate the two endpoints of each edge. Describe an algorithm that builds an edge list for any
mesh. It traverses each face of the mesh, noting each edge as it is found, but adding it only if that edge is not
already on the list.
Note that it is usually impossible to build a face list from an edge list and a vertex list. Figure 6.86 shows the
classic example. From an edge list alone there is no way to tell where the faces are: Even a wireframe model for a
cube could be a closed box or an open one. A face list has more information, therefore, than does an edge list.

Figure 6.86. An ambiguous object.

6.7.8. Case Study 6.8. Vaulted Ceilings.
(Level of Effort: III) Many classic buildings have arched ceilings or roofs shaped as a vault. Figure 6.87a shows a
domical vault [fleming66] built on a square base. Four webs, each a ruled surface, rise in a circular sweep to meet
at the peak. The arch shape is based on an ogee arch, as described in Figure 3.88. Part b shows a domical vault built on
an octagon, having eight webs. This arch shape is based on the pointed arch described in Figure 3.87. Write a function
that creates a mesh model for a domical vault built on a cube, and on a octagon.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 59

Figure 6.87. Examples of vaulted ceilings.

6.7.9. Case study 6.9. On Platonic Solids.
(Level of Effort: II) Create files (in the format described in Case Study 6.1) for each of the Platonic solids.
Experiment by reading in each file into an application and drawing the associated object.
6.7.10. Case Study 6.10. On Archimedian solids
(Level of Effort: II) Pictures of the 13 Archimedian solids may be found in many sources (e.g. [kappraff91],
[wenninger71]). Each can be formed by truncating one of the Platonic solids as described in Section 6.3.3.
Create files (in the format described in Case Study 6.1) for each of the following polyhedra:
a). the Buckyball;
b). the truncated cube;
c). the cuboctahedron. (The cube is truncated: new vertices are formed from the midpoints of each of its
edges).
Test your files by reading each object into an application and drawing it.
6.7.11. Case Study 6.11. Algebraic form for the Quadric Surfaces.
(Level of Effort: I) The implicit forms for the quadric surfaces share a compact and useful form based on
a matrix. For example, the implicit form F(x, y, z) of the generic sphere is given in Figure 6.64 as x2 + y2
+ z2 -1. By inspection this can be written as the quadratic form:

F ( x , y , z) =

1
0
( x , y , z,1)
0

0

0 0
1 0
0 1

0 x
0 y

(6.51)

0 z

1 1

0 0
T
or more compactly, using the homogeneous representation of point (x, y, z) given by P = (x, y, z, 1):

F( x, y, z ) = P T Rsphere P

(6.52)

where Rsphere is the 4 by 4 matrix displayed in Equation 6.51. The point (x, y, z) is on the ellipsoid whenever this function evaluates to 0. The implicit forms for the other quadric surfaces all have the same
form; only the matrix R is different. For instance, the matrices for the hyperboloid of two sheets and the
elliptic paraboloid are given by:

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 60

1
0
=
0
0

Rhyperboloid 2

1 0 0
0 0 0
0 1 0
1 0 0
, RellipticParab =
0 0 0
0 1 0
0 0 12
0 0 1

.

0
0
0

1
2

(6.53)

a). What the are matrices for the remaining three shapes?
Transforming quadric Surfaces.
Recall from Section 6.5.3 that when an affine transformation with matrix M is applied to a surface with
implicit function F(P), the transformed surface has implicit function F(M-1P).
b). Show that when a quadric surface is transformed, its implicit function becomes:
G( P) = ( M 1 P) T R( M 1 P) which is easily manipulated (as in Appendix 2) into the form

G( P) = P T ( M T RM 1 ) P .
Thus the transformed surface is also a quadric surface with a different defining matrix. This matrix depends both on the original shape of the quadric and on the transformation.
For example, to convert the generic sphere into an ellipsoid that extends from -a to a in x, -b to b in y,
and -c to c in z, use the scaling matrix:

a
0
M =
0
0

0 0
b 0
0 c
0 0

0
0
0
1

c). Find M-1 and show that the matrix for the ellipsoid is:

M T RM 1

1
a 2
= 0

00

0
1
b2
0
0

0
0
1
c2
0

0

0

1
0

and write the ellipsoids implicit form.

d). Find the defining matrix for an elliptic cone that has been scaled by 2 in the x-direction, by 3 in the
y-direction, and then rotated through 300 about the y-axis.
e). Show that the matrix R in the implicit function F(P)=PTRP for a quadric surface can always be taken
to be symmetric. (Hint: write R as the sum of a symmetric and an antisymmetric part, and show that the
antisymmetric part has no effect on the surface shape.
6.7.12. Case Study 6.12. Superquadric Scenes.
(Level of Effort: III) Write an application that can create generic superquadrics with user-selected bulge
factors, and which place several of them in the scene at different sizes and orientations.
6.7.13. Case Study 6.13. Drawing Smooth Parametric Surfaces.
(Level of Effort III) Develop a function that creates a mesh model of any well-behaved smooth surface given
by P(u, v) = (X(u, v), Y(u, v), Z(u, v)). It samples the surface at numValuesU uniformly spaced values in u
between uMin and uMax, and at numValuesV values in v between vMin and vMax. The functions X(), Y(),
and Z(), as well as the normal component functions nx(), ny(), and nz(), are hard-wired into the routine. It
Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 61

builds the vertex list and normal list based on these sample values, and creates a face list consisting of quadrilaterals. The only difficult part is keeping straight the indices of the vertices for each face in the face list. The
suggested skeleton shown in Figure 6.86 may prove helpful.
Apply this function to building an interesting surface of revolution, and a height field.
void Mesh:: makeSurfaceMesh()
{
int i, j, numValsU = numValsV = 40;// set these
double u, v, uMin = vMin = -10.0, uMax = vMax = 10.0;
double delU = (uMax - uMin)/(numValsU - 1);
double delV = (vMax - vMin)/(numValsV - 1);
numVerts =
numFaces =
numNorms =
pt
= new
face = new
norm = new

numValsU * numValsV + 1; // total # of vertices

(numValsU -1) * (numValsV - 1) ; // # of faces
numVerts; // for smooth shading one normal per vertex
Point3[numVerts]; assert(pt
!= NULL); // make space
Face[numFaces];
assert(face != NULL);
Vector3[numNorms]; assert(norm != NULL);

for(i = 0, u = uMin; i < numValsU; i++, u += delU)

for(j = 0, v = vMin; j < numValsV; j++, v += delV)
{
int whichVert = i * numValsV + j; //index of the vertex and normal
// set this vertex: use functions X, Y, and Z
pt[whichVert].set(X(u, v),Y(u, v),Z(u, v));
// set the normal at this vertex: use functions nx, ny, nz
norm[whichVert].set(nx(u, v), ny(u, v), nz(u, v));
normalize(norm[whichVert]);
// make quadrilateral
if(i > 0 && j > 0) // when to compute next face
{
int whichFace =(i - 1) * (numValsV - 1) + (j - 1);
face[whichFace].vert = new VertexID[4];
assert(face[whichFace].vert != NULL);
face[whichFace].nVerts = 4;
face[whichFace].vert[0].vertIndex = // same as norm index
face[whichFace].vert[0].normIndex = whichVert;
face[whichFace].vert[1].vertIndex =
face[whichFace].vert[1].normIndex = whichVert - 1;
face[whichFace].vert[2].vertIndex =
face[whichFace].vert[2].normIndex = whichVert - numValsV - 1;
face[whichFace].vert[3].vertIndex =
face[whichFace].vert[3].normIndex = whichVert - numValsV;
}
}
}
Figure 6.88. Skeleton of a mesh creation routine for a smooth surface.
6.7.14. Case Study 11.14. Taper, Twist, Bend, and Squash it.
(Level of Effort III) It is useful to have a method for deforming a 3D object in a controlled way. For instance, in an animation a bouncing rubber ball is seen to deform into a squished version as it hits the floor,
then to regain its spherical shape as it bounces up again. Or a mound of jello bends and wiggles, or a flag
waves in the breeze. In cases like these it is important to have the deformations look natural, following
rules of physics that take into account conservation of mass, elasticity, and the like. Physically-based
modeling, that attempts to mirror how actual objects behave under various forces is a large and fascinating
subject, described in many books and articles as in [watt92,bloomenthal97].
You can also produce purely geometrical deformations [barr84], chosen by the designer for their visual
effect. For instance, it is straightforward to taper an object along an axis, as suggested in Figure 6.89.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 62

Figure 6.89. The pawn before and after tapering.

This is achieved by scaling all points in the x and y dimensions by amounts that vary with z, according to
some profile function, say g(z). This defines a (non-affine) transformation that can be written as a scaling
matrix
.

g ( z)
M = 0
0

0
0
g(z) 0
0
1

(6.54)

If the undeformed surface has parametric representation P(u, v) = (X(u, v), Y(u, v), Z(u, v)) then this deformation converts it to
P(u, v) = (X(u, v)g(Z(u, v)), Y(u, v) g(Z(u, v)), Z(u, v))

(6.55)

For Figure 6.89, the mesh for the pawn was first created, and then each mesh vertex (x, y, z) was altered
to (xF, yF, z), where F is 1 - 0.04 * (z + 6) (note: the pawn extends from 0 to -12 in z).
Another useful deformation is twisting. To twist about the z-axis, for instance, rotate all points on the
object about the z-axis by an angle that depends on z, using the matrix

cos( g(z))
M = sin( g ( z ))
0

sin( g ( z)) 0
cos( g ( z )) 0
0
1

(6.56)

Figure 6.90 shows the pawn after a linearly increasing twist is applied. The pawn is a surface of revolution about the z-axis, so it doesnt make much sense to twist it about the z-axis. Instead the twist here is
about the y-axis, with g( z ) = 0.02 z + 6 .

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 63

Figure 6.90. The pawn after twisting.

Apply these deformations to several mesh models, including the torus. (Note that you cannot use
OpenGLs modelview matrix to perform this deformation, since the transformation here is not affine.
The vertices in the actual vertex list must be transformed. Bending is another deformation treated by
Barr. Refer to his paper [barr84], and experiment with the bending deformation as well.

6.8. For Further Reading

A number of books are available that treat the definition and generation of surfaces and solids. Rogers and
Adams MATHEMATICAL ELEMENTS FOR COMPUTER GRAPHICS [rogers90] provides a clear introduction to curves and surfaces, as does Faux and Pratts COMPUTATIONAL GEOMETRY FOR DESIGN
AND MANUFACTURE [Faux79]. Grays MODERN DIFFERENTIAL GEOMETRY OF CURVES AND
SURFACES WITH MATHEMATICA [gray93] offers a rigorous mathematical treatment of curve and surface
shapes, and provides code in Mathematica for drawing them. Mortensons GEOMETRIC MODELING
[mortenson85] also provides an excellent discussion of solid modeling used in the CAD industry.

Hill - ECE660:Computer Graphics Chapter 6

10/23/99

page 64

(for ECE660, Fall, 1999)

CHAPTER 7 Three-Dimensional Viewing

I am a camera with its shutter open, quite passive, recording, not thinking.
Christopher Isherwood, A Berlin Diary

Goals of the Chapter

To develop tools for creating and manipulating a camera that produces pictures of a 3D scene.
To see how to fly a camera through a scene interactively, and to make animations.
To learn the mathematics that describe various kinds of projections.
To see how each operation in the OpenGL graphics pipeline operates, and why it is used.
To build a powerful clipping algorithm for 3D objects.
To devise a means for producing stereo views of objects.
Preview
Section 7.1 provides an overview of the additional tools that are needed to build an application that lets a
camera fly through a scene. Section 7.2 defines a camera that produces perspective views, and shows
how to make such a camera using OpenGL. It introduces aviation terminology that helps to describe ways
to manipulate a camera. It develops some of the mathematics needed to describe a cameras orientation
through a matrix. Section 7.3 develops the Camera class to encapsulate information about a camera, and
develops methods that create and adjust a camera in an application.
Section 7.4 examines the geometric nature of perspective projections, and develops mathematical tools to
describe perspective. It shows how to incorporate perspective projections in the graphics pipeline, and
describes how OpenGL does it. An additional property of homogeneous coordinates is introduced to
facilitate this. The section also develops a powerful clipping algorithm that operates in homogeneous
coordinate space, and shows how its efficiency is a result of proper transformations applied to points
before clipping begins. Code for the clipper is given for those programmers who wish to develop their
own graphics pipeline.
Section 7.5 shows how to produce stereo views of a scene in order to make them more intelligible.
Section 7.6 develops a taxonomy of the many kinds of projections used in art, architecture, and
engineering, and shows how to produce each kind of projection in a program. The chapter closes with a
number of Case Studies for developing applications that test the techniques discussed.

7.1 Introduction.
We are already in a position to create pictures of elaborate 3D objects. As we saw in Chapter 5, OpenGL
provides tools for establishing a camera in the scene, for projecting the scene onto the cameras
viewplane, and for rendering the projection in the viewport. So far our camera only produces parallel
projections. In Chapter 6 we described several classes of interesting 3D shapes that can be used to model
the objects we want in a scene, and through the Mesh class we have ways of drawing any of them with
appropriate shading.
So whats left to do? For greater realism we want to create and control a camera that produces
perspective projections. We also need ways to take more control of the cameras position and orientation,
so that the user can fly the camera through the scene in an animation. This requires developing more
controls than OpenGL provides. We also need to achieve precise control over the cameras view volume,
which is determined in the perspective case as it was when forming parallel projections: by a certain
matrix. This requires a deeper use of homogeneous coordinates than we have used so far, so we develop
the mathematics of perspective projections from the beginning, and see how they are incorporated in the
OpenGL graphics pipeline. We also describe how clipping is done against the cameras view volume,
which again requires some detailed working with homogeneous coordinates. So we finally see how it is
all done, from start to finish! This also provides the underlying theory for those programmers who must
develop 3D graphics software without the benefit of OpenGL.

7.2. The Camera Revisited.

It adds a precious seeing to the eye.

Chapter

3D viewing

November 23, 1999

page 1

William Shakespeare, Love's Labours Lost

In Chapter 5 we used a camera that produces parallel projections. Its view volume is a parallelepiped
bounded by six walls, including a near plane and a far plane. OpenGL also supports a camera that creates
perspective views of 3D scenes. It is similar in many ways to the camera used before, except that its view
volume has a different shape.
Figure 7.1 shows its general form. It has an eye positioned at some point in space, and its view volume is
a portion of a rectangular pyramid, whose apex is at the eye. The opening of the pyramid is set by the
viewangle (see part b of the figure). Two planes are defined perpendicular to the axis of the pyramid:
the near plane and the far plane. Where these planes intersect the pyramid they form rectangular
windows. The windows have a certain aspect ratio, which can be set in a program. OpenGL clips off any
parts of the scene that lie outside the view volume. Points lying inside the view volume are projected
onto the view plane to a corresponding point P as suggested in part c. (We shall see that it doesnt
matter which plane one uses as the view plane, so for now take it to be the near plane.) With a
perspective projection the point P is determined by finding where a line from the eye to P intersects the
view plane. (Contrast this with how a parallel projection operates.)
a).
b).
c).

Figure 7.1. A camera to produce perspective views of a scene.

Finally, the image formed on the view plane is mapped into the viewport as shown in part c, and becomes
visible on the display device.
7.2.1. Setting the View Volume.
Figure 7.2 shows the camera in its default position, with the eye at the origin and the axis of the pyramid
aligned with the z-axis. The eye is looking down the negative z-axis.

Figure 7.2. the camera in its default position.

OpenGL provides a simple way to set the view volume in a program. Recall that the shape of the
cameras view volume is encoded in the projection matrix that appears in the graphics pipeline. The
projection matrix is set up using the function gluPerspective() with four parameters. The sequence
to use is:
glMatrixMode(GL_PROJECTION); // make the projection matrix current
glLoadIdentity();
// start with a unit matrix
gluPerspective(viewAngle, aspectRatio, N, F); // load the appropriate
values
The parameter viewAngle, shown as in the figure, is given in degrees, and sets the angle between the
top and bottom walls of the pyramid. aspectRatio sets the aspect ratio of any window parallel to the
xy-plane. The value N is the distance from the eye to the near plane, and F is the distance from the eye to
the far plane. N and F should be positive. For example, gluPerspective(60.0, 1.5, 0.3,
o
50.0) establishes the view volume to have a vertical opening of 60 , with a window that has an aspect

Chapter

3D viewing

November 23, 1999

page 2

ratio of 1.5. The near plane lies at z = -0.3 and the far plane lies at z = -50.0. We see later exactly what
values this function places in the projection matrix.
7.2.2. Positioning and pointing the camera.
In order to obtain the desired view of a scene, we move the camera away from its default position shown
in Figure 7.2, and aim it in a particular direction. We do this by performing a rotation and a translation,
and these transformations become part of the modelview matrix, as we discussed in Section 5.6.
We set up the cameras position and orientation in exactly the same way as we did for the parallelprojection camera. (The only difference between a parallel- and perspective-projection camera resides in
the projection matrix, which determines the shape of the view volume.) The simplest function to use is
again gluLookAt(), using the sequence
glMatrixMode(GL_MODELVIEW); // make the modelview matrix current
glLoadIdentity();
// start with a unit matrix
gluLookAt(eye.x, eye.y, eye.z, look.x, look.y, look.z, up.x, up.y,
up.z);
As before this moves the camera so its eye resides at point eye, and it looks towards the point of interest
look. The upward direction is generally suggested by the vector up, which is most often set simply to
(0, 1, 0). We took these parameters and the whole process of setting the camera pretty much for granted
in Chapter 5. In this chapter we will probe deeper, both to see how it is done and to take finer control
over setting the camera. We also develop tools to make relative changes to the cameras direction, such
as rotating it slightly to the left, tilting it up, or sliding it forward.
The General camera with arbitrary orientation and position.
A camera can have any position in the scene, and any orientation. Imagine a transformation that picks up
the camera of Figure 7.2 and moves it somewhere in space, then rotates it around to aim it as desired. We
need a way to describe this precisely, and to determine what the resulting modelview matrix will be.
It will serve us well to attach an explicit coordinate system to the camera, as suggested by Figure 7.3.
This coordinate system has its origin at the eye, and has three axes, usually called the u-, v-, and n- axes,
that define its orientation. The axes are pointed in directions given by the vectors u, v, and n as shown in
the figure. Because the camera by default looks down the negative z-axis, we say in general that the
camera looks down the negative n-axis, in the direction -n. The direction u points off to the right of
the camera, and direction v points upward. Think of the u-, v-, and n-axes as clones of the x-,y-, and
z-axes of Figure 7.2, that are moved and rotated as we move the camera into position.

Figure 7.3. Attaching a coordinate system to the camera.

Position is easy to describe, but orientation is difficult. It helps to specify orientation using the flying
terms pitch, heading, yaw, and roll, as suggested in Figure 7.4. The pitch of a plane is the angle that its
longitudinal axis (running from tail to nose and having direction -n) makes with the horizontal plane. A
plane rolls by rotating about this longitudinal axis; its roll is the amount of this rotation relative to the
horizontal. A planes heading is the direction in which it is headed. (Other terms are azimuth and
bearing.) To find the heading and pitch given n, simply express -n in spherical coordinates, as shown in
Figure 7.5. (See Appendix 2 for a review of spherical coordinates.) The vector -n has longitude and
latitude given by angles and , respectively. The heading of a plane is given by the longitude of -n, and
the pitch is given by the latitude of -n. Formulas for roll, pitch, and heading in terms of the vectors u and
n are developed in the exercises.

Chapter

3D viewing

November 23, 1999

page 3

a).pitch

b).roll

c).yaw
v

n
Figure 7.4. A planes orientation relative to the world.

Figure 7.5. The heading and pitch of a plane.

Pitch and roll are both nouns and verbs: when used as verbs they describe a change in the planes
orientation. You can say a plane pitches up when it increases its pitch (rotates about its u-axis), and
that it rolls when it rotates about its n-axis. The common term for changing heading is yaw: to yaw left
or right it rotates about its v-axis.
These terms can be used with a camera as well. Figure 7.6a shows a camera with the same coordinate
system attached: it has u, v, and n- axes, and its origin is at position eye. The camera in part b has some
non-zero roll, whereas the one in part c has zero roll. We most often set a camera to have zero roll, and
call it a no-roll camera. The u-axis of a no-roll camera is horizontal: that is, perpendicular to the y-axis
of the world. Note that a no-roll camera can still have an arbitrary n direction, so it can have any pitch or
heading.
How do we control the roll, pitch, and heading of a camera? gluLookAt() is handy for setting up an
initial camera, since we usually have a good idea of how to choose eye and look. But its harder to
visualize how to choose up to obtain a certain roll, and its hard to make later relative adjustments to the
camera using only gluLookAt(). (gluLookAt() works with Cartesian coordinates, whereas orientation
deals with angles and rotations about axes.) OpenGL doesnt give direct access to the u, v, and n
directions, so well maintain them ourselves in a program. This will make it much easier to describe and
adjust the camera.
a). Cam era orie ntation

b). with roll

c). n o- roll

v
v
u

n
Figure 7.6. Various camera orientations.
What gluLookAt() does: some mathematical underpinnings.
What then are the directions u, v, and n when we execute gluLookAt() with given values for eye, look,
and up? Lets see exactly what gluLookAt() does, and why it does it.
As shown in Figure 7.7a, we are given the locations of eye and look, and the up direction. We
immediately know that n must be parallel to the vector eye - look, as shown in Figure 7.7b, so we set
n = eye - look. (Well normalize this and the other vectors later as necessary.)

Chapter

3D viewing

November 23, 1999

page 4

Figure 7.7. Building the vectors u, v, and n.

We now need to find u and v that are perpendicular to n and to each other. The u direction points off to
the side of a camera, so it is (fairly) natural to make it perpendicular to up, which the user has said is the
upward direction. This is the assumption gluLookAt() makes in any case, and so the direction u is
made perpendicular to n and up. An excellent way to build a vector that is perpendicular to two given
vectors is to form their cross product, so we set u = up n . (The user should not choose an up
direction that is parallel to n, as then u would have zero length - why?) We choose u = up n rather
than n up so that u will point to the right as we look along -n.
With u and n in hand it is easy to form v: it must be perpendicular to both u and v so use a cross product
again: v = n u . Notice that v will usually not be aligned with up: v must be aimed perpendicular to n,
whereas the user provides up as a suggestion of upwardness, and the only property of it that is used is
its cross product with n.
Summarizing: given eye, look, and up, we form

n = eye look
u = up n
v = nu

(7.1)

and then normalize all three to unit length.

Note how this plays out for the common case where up = (0,1,0). Convince yourself that in this case u =
(nz, 0, -nx) and v = (-nxny, nx2 + nz2, -nzny). Notice that u does indeed have a y-component of 0, so it is
horizontal. Further, v has a positive y-component, so it is pointed more or less upward.
Example 7.2.1. Find the camera coordinate system. Consider a camera with eye = (4, 4, 4) that looks
down on a look-at point look = (0, 1, 0). Further suppose that up is initially set to (0, 1, 0). Find u, v,
and n. Repeat for up = (2, 1, 0).
Solution: From Equation 7.1 we find: u = (4, 0, -4) , v = (-12, 32, -12) , n = (4, 3, 4), which are easily
normalized to unit length. (Sketch this situation.) Note that u is indeed horizontal. Check that these are
mutually perpendicular. For the case of up = (2, 1, 0) (try to visualize this camera before working out the
arithmetic), u = (4, -8, 2), v = (38, 8, -44), and n = (4, 3, 4). Sketch this situation. Check that these
vectors are mutually perpendicular.
Example 7.2.2. Building intuition with cameras. To assist in developing geometric intuition when
setting up a camera, Figure 7.8 shows two example cameras - each depicted as a coordinate system with a
view volume - positioned above the world coordinate system, which is made more visible by grids drawn
in the xz-plane. The view volume of both cameras has an aspect ratio of 2. One camera is set with eye =
(-2, 2, 0), look = (0,0,0), and up = (0,1,0). For this camera we find from Equation 7.1 that n = (-2, 2, 0), u
= (0,0,2), and v = (4, 4, 0). The figure shows these vectors (v is drawn the darkest) as well as the up
vector. The second camera uses eye = (2,2,0), look = (0,0,0), and up = (0,0,1). In this case u = (-2, -2, 0)
and v = (0, 0, 8). The direction v is parallel to up here. Note that this camera appears to be on its side:
(Check that all of these vectors appear drawn in the proper directions.)

Chapter

3D viewing

November 23, 1999

page 5

Figure 7.8. Two example settings of the camera.

Finally, we want to see what values gluLookAt() places in the modelview matrix. From Chapter 5 we
know that the modelview matrix is the product of two matrices, the matrix V that accounts for the
transformation of world points into camera coordinates, and the matrix M that embodies all of the
modeling transformations applied to points. gluLookAt() builds the V matrix and postmultiplies the
current matrix by it. Because the job of the V matrix is to convert world coordinates to camera
coordinates, it must transform the cameras coordinate system into the generic position for the camera as
shown in Figure 7.9. This means it must transform eye into the origin, u into the vector i, v into j, and n
into k. There are several ways to derive what V must be, but its easiest to check that the following
matrix does the trick:

Figure 7.9. The transformation which gluLookAt() sets up.

u
v
V=
n
0

x
x
x

uy
vy
ny
0

uz
vz
nz
0

dx
dy
dz
1

(7.2)

where ( d x , d y , d z ) = ( eye u, eye v, eye n) 1. Check that in fact

eye 0
eye = 0
V
eye 0
1 1
x
y
z

1 A technicality: since its not legal to dot a point and a vector, eye should be replaced here by the vector
(eye - (0,0,0)).

Chapter

3D viewing

November 23, 1999

page 6

u 1
u 0
V =
u 0
0 0

as desired, where we have extended point eye to homogeneous coordinates. Also check that
x
y
z

and that V maps v into (0,1,0,0)T and maps n into (0,0,1,0)T. The matrix V is created by gluLookAt()
and is postmultiplied with the current matrix. We will have occasion to do this same operation later when
we maintain our own camera in a program.
Practice Exercises.
7.2.1. Finding roll, pitch, and heading given vectors u, v, and n. Suppose a camera is based on a
coordinate system with axes in the directions u, v, and n, all unit vectors. The heading and pitch of the
camera is found by representing -n in spherical coordinates. Using Appendix 2, show that

heading = arctan( n z , n x )
pitch = sin 1 ( n y )
Further, the roll of the camera is the angle its u-axis makes with the horizontal. To find it, construct a
vector b that is horizontal and lies in the uv-plane. Show that b = j n has these properties. Show that
the angle between b and u is given by

roll = cos 1

u n
n
x

z
2

uz n x
+ nz 2

7.2.2. Using up sets v to a best approximation to up. Show that using up as in Equation 7.1 to set u
and v is equivalent to making v the closest vector to up that is perpendicular to vector n. Use these steps:
a). Show that v = n ( up n) ;
b). Use a property of the triple vector product, that says a (b c) = (a c)b (a b)c .
c). Show that v is therefore the projection of up onto the plane with normal n (see Chapter 4), and
therefore is the closest vector in this plane to up.

7.3 Building a Camera in a Program.

It is as interesting and as difficult to say a thing well as to paint it.
Vincent van Gogh
In order to have fine control over camera movements, we create and manipulate our own camera in a
program. After each change to this camera is made, the camera tells OpenGL what the new camera is.
We create a Camera class that knows how to do all the things a camera does. Its very simple and the
payoff is high. In a program we create a Camera object called, say, cam, and adjust it with functions
such as:
cam.set(eye, look, up);
// initialize the camera - similar to
gluLookAt()
cam.slide(-1,0,-2); // slide the camera
forward and to the left
o
cam.roll(30); // roll it through 30
o
cam.yaw(20); // yaw it through 20
etc.
Figure 7.10 shows the basic definition of the Camera class. It contains fields for the eye and the
directions u, v, and n. (Point3 and Vector3 are the basic data types defined in Appendix 3.) It also
has fields that describe the shape of the view volume: viewAngle, aspect, nearDist, and
farDist.
class Camera{

Chapter

3D viewing

November 23, 1999

page 7

The first question is then, what are x* and y*? Its simplest to use similar triangles, and say x* is in the
same ratio to Px as the distance N is to the distance |Pz|. Or since Pz is negative, we can say

x*
N
=
Pz
Px
or x* = NPx/(-Pz). Similarly y* = NPy/(-Pz). So we have that P projects to the point on the viewplane:

( x*, y*) = N

Py
Px
,N
Pz
Pz

(the projection of P)

(7.4)

An alternate (analytical) method for arriving at this result is given in the exercises.
Example 7.4.1: Where on the viewplane does P = (1, 0.5, -1.5) lie for the camera having a near plane at
N = 1? Solution: Direct use of Equation 7.4 yields (x*, y*) = (0.666, 0.333).
We can make some preliminary observations about how points are projected.
1). Note the denominator term -Pz. It is larger for more remote points (those farther along the negative zaxis), which reduces the values of x* and y*. This introduces perspective foreshortening, and makes
remote parts of an object appear smaller than nearer parts.
2). Denominators have a nasty way of evaluating to zero, and Pz becomes 0 when P lies in the same
plane as the eye: the z = 0 plane. Normally we use clipping to remove such offending points before
trying to project them.
3). If P lies behind the eye there is a reversal in sign of Pz, which causes further trouble, as we see later.
These points, too, are usually removed by clipping.
4). The effect of the near plane distance N is simply to scale the picture (both x* and y* are proportional
to N). So if we choose some other plane (still parallel to the near plane) as the view plane onto which to
project pictures, the projection will differ only in size with the projection onto the near plane. Since we
ultimately map this projection to a viewport of a fixed size on the display, the size of the projected image
makes no difference. This shows that any viewplane parallel to the near plane would work just as well, so
we might as well use the near plane itself.
Chapter

3D viewing

November 23, 1999

page 12

5). Straight lines project to straight lines. Figure 7.17 provides the simplest proof. Consider the line in 3D
space between points A and B. A projects to A and B projects to B. But do points between A and B
project to points directly between A and B? Yes: just consider the plane formed by A, B and the origin.
Since any two planes intersect in a straight line, this plane intersects the near plane in a straight line.
Thus line segment AB projects to line segment AB.

Figure 7.17. Proof that a straight line projects to a straight line.

Example 7.4.2. Three projections of the barn.
A lot of intuition can be acquired by seeing how a simple object is viewed by different cameras. Here we
examine how the edges of the barn defined in Chapter 6 and repeated in Figure 7.18 are projected onto
three cameras. The barn has 10 vertices, 15 edges and seven faces.

Figure 7.18. The basic barn revisited.

View #1: We first set the cameras eye at (0, 0, 3) and have it look down the negative z-axis, with u =
(1,0,0) and n = (-1, 0, 0). We will set the near plane at distance 1 from the eye. (The near plane happens
therefore to coincide with the front of the barn.) In terms of camera coordinates all points on the front
wall of the barn have Pz = -1 and those on the back wall have Pz = -2. So from Equation 7.4 any point (Px,
Py, Pz) on the front wall projects to:
P = (Px, Py)

{projection of a point on the front wall}

and any point on the back wall projects to

P = (Px /2, Py / 2).

{projection of a point on the back wall}

The foreshortening factor is two for those points on the back wall. Figure 7.19a shows the projection of
the barn for this view. Note that edges on the rear wall project at half their true length. Also note that
edges of the barn that are actually parallel in 3D need not project as parallel. (We shall see that parallel
edges that are parallel to the viewplane do project as parallel, but parallel edges that are not parallel to
the viewplane are not parallel: they recede to a vanishing point.)

Figure 7.19. Projections of the barn for views #1 and #2.

View #2: Here the camera has been slid over so that eye = (0.5, 0, 2), but u, and n are the same as in view
#1. Figure 7.19b shows the projection.
View #3 Here we use the camera with Eye = (2, 5, 2) and look = (0,0,0), resulting in Figure 7.20. The
world axes have been added as a guide. This shows the barn from an informative point of view. From a
wireframe view it is difficult to discern which faces are where.

Figure 7.20. A third view of the barn.

Practice Exercises.
7.4.1. Sketch a Cube in Perspective. Draw (by hand) the perspective view of a cube C (axis-aligned and
centered at the origin, with sides of length 2) when the eye is at E = 5 on the z-axis. Repeat when C is
shifted so that its center is at (1, 1, 1).
7.4.2. Where does the ray hit the viewplane? (Dont skip this one.)
We want to derive Equation 7.4 by finding where the ray from the origin to P intersects the near plane.
Chapter

3D viewing

November 23, 1999

page 13

a). Show that if this ray is at the origin at t = 0 and at P at time t = 1, then it has parametric representation
r(t) = Pt.
b). Show that it hits the near plane at t = N/(-Pz);
c). Show that the hit point is (x*, y*) = (NPx/(-Pz), NPy/(-Pz)).
7.4.2. Perspective Projection of a Line.
We develop here some interesting properties of perspective projections by examining how straight lines
project.
1). Lines that are parallel in 3D project to lines, but these lines arent necessary parallel. If not parallel,
they meet at some vanishing point.
2). Lines that pass behind the eye of the camera cause a catastrophic passage through infinity. (Such
lines should be clipped off.)
3). Perspective projections usually produce geometrically realistic pictures. But realism is strained for
very long lines parallel to the viewplane.
1). Projecting Parallel Lines.
We suppose the line in 3D passes (using camera coordinates) through point A = (Ax, Ay, Az) and has
direction vector c = (cx, cy, cz). It therefore has parametric form P(t) = A + c t. Substituting this form in
Equation 7.4 yields the parametric form for the projection of this line:

p(t ) = N

Ay + c y t
Ax + c x t
,N
Az c z t
Az c z t

(7.5)

(This may not look like the parametric form for a straight line, but it is. See the exercises.) Thus the point
A in 3D projects to the point p(0), and as t varies the projected point p(t) moves across the screen (in a
straight line). We can discern several important properties directly from this formula.
Suppose the line A + ct is parallel to the viewplane. Then cz = 0 and the projected line is given by:

p(t ) =

N
Ax + c x t, Ay + c y t
Az

This is the parametric form for a line with slope cy/cx. This slope does not depend on the position of the
line, only its direction c. Thus all lines in 3D with direction c will project with this slope, so their
projections are parallel. We conclude:
If two lines in 3D are parallel to each other and to the viewplane, they project to two parallel lines.
Now consider the case where the direction c is not parallel to the viewplane. For convenience suppose cz
< 0, so that as t increases the line recedes further and further from the eye. At very large values of t,
Equation 7.5 becomes:

p() = N

cy
cx
,N
cz
cz

(the vanishing point for the line)

(7.6)

This is called the vanishing point for this line: its the point towards which the projected line moves as
t gets larger and larger. Notice that it depends only on the direction c of the line and not its position
(which is embodied in A). Thus all parallel lines share the same vanishing point.
In particular, these lines project to lines that are not parallel.
Figure 7.21a makes this more vivid for the example of a cube. Several edges of the cube are parallel:
there are those that are horizontal, those that are vertical, and those that recede from the eye. This picture

Chapter

3D viewing

November 23, 1999

page 14

was made with the camera oriented so that its near plane was parallel to the front face of the cube. Thus
in camera coordinates the z-component of c for the horizontal and vertical edges is 0. The horizontal
edges therefore project to parallel lines, and so do the vertical edges. The receding edges, however, are
not parallel to the view plane, and hence converge onto a vanishing point (VP). Artists often set up
drawings of objects this way, choosing the vanishing point and sketching parallel lines as pointing at the
VP. We shall see more on vanishing points as we proceed.
VP

d
Figure 7.21. The vanishing point for parallel lines.
Figure 7.22 suggests what a vanishing point is geometrically. Looking down onto the cameras xz-plane
from above, we see the eye viewing various points on the line AB. A projects to A, B projects to B, etc.
Very remote points on the line project to VP as shown. The point VP is situated so that the line from the
eye through VP is parallel to AB (why?).

Figure 7.22. The geometry of a vanishing point.

2). Lines that Pass Behind the Eye.
We saw earlier that trying to project a point that lies in the plane of the eye (z = 0 in eye coordinates)
results in a denominator of 0, and would surely spell trouble if we try to project it. We now examine the
projection of a line segment where one endpoint lies in front of the eye, and one endpoint lies behind.
Figure 7.23 again looks down on the camera from above. Point A lies in front of the eye and projects to
A in a very reasonable manner. Point B, on the other hand, lies behind the eye, and projects to B, which
seems to end up on the wrong side of the viewplane! Consider a point C that moves from A to B, and
sketch how its projection moves. As C moves back towards the plane of the eye, its projection slides
further and further along the viewplane to the right. As C approaches the plane of the eye its projection
spurts off to infinity, and as C moves behind the eye its projection reappears from far off to the left on the
viewplane! You might say that the projection has wrapped around infinity and come back from the
opposite direction [blinn96]. If we tried to draw such a line there would most likely be chaos. Normally
all parts of the line closer to the eye than the near plane are clipped off before the projection is attempted.

Figure 7.23. Projecting the line segment AB, with B behind the eye.
Example 7.4.3. The Classic Horizontal Plane in Perspective.
A good way to gain insight into vanishing points is to view a grid of lines in perspective, as in Figure 7.24.
Grid lines here lie in the xz-plane, and are spaced 1 unit apart. The eye is perched 1 unit above the xz-plane,
at (0, 1, 0), and looks along -n, where n = (0,0,1). As usual we take up = (0,1,0). N is chosen to be 1.

Chapter

3D viewing

November 23, 1999

page 15

Figure 7.24. Viewing a horizontal grid on the xz-plane.

The grid lines of constant x have parametric form in eye coordinates of (i, 0, t), where i = ..,-2, -1, 0, 1,
2, ... and t varies from 0 to . By Equation 7.4 the i-th line projects to (-i/t, 2/t), which is a line through
the vanishing point (0, 0), so all of these lines converge on the same vanishing point, as expected.
The grid lines of constant z are given by (t, 0, -i), where i = 1, 2, ..N for some N., and t varies from -
to . These project to (-t/i, -2/i), which appear as horizontal straight lines (check this). Their projections
are parallel since the gridlines themselves are parallel to the viewplane. The more remote ones (larger
values of i) lie closer together, providing a vivid example of perspective foreshortening. Many of the
remote contours are not drawn here, as they become so crowded they cannot be drawn clearly. The
horizon consists of all the contours where z is very large and negative; it is positioned at y = 0.
3). The anomaly of viewing long parallel lines.
Perspective projections seem to be a reasonable model for the way we see. But there are some anomalies,
mainly because our eyes do not have planar view screens. The problem occurs for very long objects.
Consider an experiment, for example, where you look up at a pair of parallel telephone wires, as
suggested in Figure 7.25a.
a). viewing parallel
telephone wires

b). A perspective
projection shows

c). What your eye sees

Figure 7.25. Viewing very long parallel wires. (use old 12.14).
For the perspective view, if we orient the viewplane to be parallel to the wires, we know the image will
show two straight and parallel lines (part b). But what you see is quite different. The wires appear curved
as they converge to vanishing points in both directions (part c)! In Practice this anomaly is barely

Chapter

3D viewing

November 23, 1999

page 16

visible because the window or your eye limits the field of view to a reasonable region. (To see different
parts of the wires you have to roll your eyes up and down, which of course rotates your view planes.)
Practice Exercises.
7.4.3. Straight lines project as straight lines: the parametric form. Show that the parametric form in
Equation 7.5 is that of a straight line. Hint: For the x-component divide the denominator into the
numerator to get -AxN/Az + R g(t) where R depends on the x-components of A and c, but not the ycomponents, and g(t) is some function of t that depends on neither the x nor y-components. Repeat for the
y-component, obtaining -AyN/Az + Sg(t) with similar properties. Argue why this is the parametric
representation of a straight line, (albeit one for which the point does not move with constant speed as t
varies).
7.4.4. Derive results for horizontal grids. Derive the parametric forms for the projected grid lines in
Example 7.4.3.
7.4.3. Incorporating Perspective in the Graphics Pipeline.
Only a fool tests the depth of the river with both feet.
Paul Cezanne, 1925
We want the graphics pipeline to project vertices of 3D objects onto the near plane, then map them to the
viewport. After passing through the modelview matrix, the vertices are represented in the cameras
coordinate system, and Equation 7.4 shows the values we need to compute for the proper projection. We
need to do clipping, and then map what survives to the viewport. But we need a little more as well.
Adding Pseudodepth.
Taking a projection discards depth information; that is, how far the point is from the eye. But we mustnt
discard this information completely, or it will be impossible to do hidden surface removal later.
The actual distance of a point P from the eye in camera coordinates is

Px 2 + Py 2 + Pz 2 , which would

be cumbersome and slow to compute for each point of interest. All we really need is some measure of
distance that tells when two points project to the same point on the near plane, which is the closer. Figure
7.26 shows points P1 and P2 that both lie on a line from the eye, and therefore project to the same point.
We must be able to test whether P1obscures P2 or vice versa. So for each point P that we project we
compute a value called the pseudodepth that provides an adequate measure of depth for P. We then say
that P projects to (x*, y*, z*), where (x*, y*) is the value provided in Equation 7.4 and z* is its
pseudodepth.

Figure 7.26. Is P1 closer than P2 or farther away?

What is a good choice for the pseudodepth function? Notice that if two points project to the same point
the farther one always has a more negative value of Pz, so we might use -Pz itself for pseudodepth. But it
will turn out to be very harmonious and efficient to choose a function with the same denominator (-Pz) as
occurs with x* and y*. So we try a function that has this denominator, and a numerator that is linear in Pz,
and say that P projects to

( x, y, z*) = N

Py aPz + b
Px
,N
,
Pz
Pz Pz

(7.7)

for some choice of the constants a and b. Although many different choices for a and b will do, we choose
them so that the pseudodepth varies between -1 and 1 (we see later why these are good choices). Since
depth increases as a point moves further down the negative z-axis, we decide that the pseudodepth is -1
when Pz = -N, and is +1 when Pz = -F. With these two conditions we can easily solve for a and b,
obtaining:

Chapter

2 FN
F+N
,b =
FN
FN

3D viewing

(7.8)

November 23, 1999

page 17

Figure 7.27 plots pseudodepth versus (-Pz). As we insisted it grows from -1 for a point on the near plane
up to +1 for a point on the far plane. As Pz approaches 0 (so that the point is just in front of the eye)
pseudodepth plummets to -. For a point just behind the eye, the pseudodepth is huge and positive. But
we will clip off points that lie closer than the near plane, so this catastrophic behavior will never be
encountered.

Figure 7.27. Pseudodepth grows as Pz becomes more negative.

Also notice that pseudodepth values bunch together as (-Pz) gets closer to F. Given the finite precision
arithmetic of a computer, this can cause a problem when you must distinguish the pseudodepth of two
points during hidden surface removal: the points have different true depths from the eye, but their
pseudodepths come out with the same value!
Note that defining pseudodepth as in Equation 7.7 causes it to become more positive as Pz becomes more
negative. This seems reasonable since depth from the eye increases as Pz moves further along the
negative z-axis.
Example 7.4.4. Pseudodepth varies slowly as (-Pz) approaches F.
Suppose N = 1 and F = 100. This produces a = -101/99 and b = -200/99, so pseudodepth is given by

pseudodepth N =1, F =100 =

101Pz + 200
99 Pz

This maps appropriately to -1 at Pz = -N, and 1 at Pz = -F. But close to -F it varies quite slowly with (-Pz).
For (-Pz ) values of 97, 98, and 99, for instance, this evaluates to 1.041028, 1.040816, and 1.040608.
A little algebra (see the exercises) shows that when N is much smaller than F as it normally will be,
pseudodepth can be approximated by

pseudodepth 1 +

2N
Pz

(7.9)

Again you see that it varies more and more slowly as (-Pz) approaches F. But its variation is increased by
using large values of N. N should be set as large as possible (but of course not so large that objects
nearest to the camera are clipped off!).
Using Homogeneous Coordinates.
Why was there consideration given to having the same denominator for each term in Equation 7.7? As we
now show, this makes it possible to represent all of the steps so far in the graphics pipeline as matrix
multiplications, offering both processing efficiency and uniformity. (Chips on some graphics cards can
multiply a point by a matrix in hardware, making this operation extremely fast!) Doing it this way will
also allow us to set things up for a highly efficient and reliable clipping step.
The new approach requires that we represent points in homogeneous coordinates. We have been doing
that anyway, since this makes it easier to transform a vertex by the modelview matrix. But we are going
to expand the notion of the homogeneous coordinate representation beyond what we have needed before
now, and therein find new power. In particular, a matrix will not only be able to perform an affine
transformation, it will be able to perform a perspective transformation.
Up to now we have said that a point P = (Px, Py, Pz) has the representation (Px, Py, Pz, 1) in homogeneous
coordinates, and that a vector v = (vx, vy, vz) has the representation (vx, vy, vz, 0 ). We have simply
appended a 1 or 0. This made it possible to use coordinate frames as a basis for representing the points
and vectors of interest, and it allowed us to represent an affine transformation by a matrix.
Now we extend the idea , and say that a point P = (Px, Py, Pz) has a whole family of homogeneous
representations (wPx, wPy, wPz, w) for any value of w except 0. For example, the point (1, 2, 3) has the
representations (1, 2, 3, 1), (2, 4, 6, 2), (0.003, 0.006, 0.009, 0.001), (-1, -2, -3, -1), and so forth. If
Chapter

3D viewing

November 23, 1999

page 18

someone hands you a point in this form, say (3, 6, 2, 3) and asks what point is it, just divide through by
the last component to get (1, 2, 2/3, 1), then discard the last component: the point in ordinary
coordinates is (1, 2, 2/3). Thus:

To convert a point from ordinary coordinates to homogeneous coordinates, append a 14;

To convert a point from homogeneous coordinates to ordinary coordinates, divide all components by
the last component, and discard the fourth component.

The additional property of being able to scale all the components of a point without changing the point is
really the basis for the name homogeneous. Up until now we have always been working with the
special case where the final component is 1.
We examine homogeneous coordinates further in the exercises, but now focus on how they operate when
transforming points. Affine transformations work fine when homogeneous coordinates are used. Recall
that the matrix for an affine transformation always has (0,0,0,1) in its fourth row. Therefore if we
multiply a point P in homogeneous representation by such a matrix M, to form MP = Q (recall Equation
5.24), as in the example

2
6
0
0

1
.5
4
0

3 1 wPx
wQx
1 4 wPy
wQy
=
2 3 wPz
wQz
0 1
w
w

the final component of Q will always be unaltered: it is still w. Therefore we can convert the Q back to
ordinary coordinates in the usual fashion.
But something new happens if we deviate from a fourth row of (0,0,0,1). Consider the important example
that has a fourth row of (0, 0, -1, 0), (which is close to what we shall later call the projection matrix):

N
0
0
0

0
N
0
0

b

0

0
0

a
1

(the projection matrix - version 1)

(7.10)

for any choices of N, a, and b. Multiply this by a point represented in homogeneous coordinates with an
arbitrary w:

N
0
0
0

0
N
0
0

wNPx
0 wPx
wNPy
0 wPy
=
w(aPz + b )
a b wPz
wPz
1 0 w
0
0

This corresponds to an ordinary point, but which one? Divide through by the fourth component and
discard it, to obtain

N P , N P , aP + b
P P P
y

and, if you wish, multiply all four components by any nonzero value.

Chapter

3D viewing

November 23, 1999

page 19

which is precisely what we need according to Equation 7.7. Thus using homogeneous coordinates allows
us to capture perspective using a matrix multiplication! To make it work we must always divide through
by the fourth component, a step which is called perspective division.
A matrix that has values other than (0,0,0,1) for its fourth row does not perform an affine transformation.
It performs a more general class of transformation called a perspective transformation. It is a
transformation not a projection. A projection reduces the dimensionality of a point, to a 3-tuple or a 2tuple, whereas a perspective transformation takes a 4-tuple and produces a 4-tuple.
Consider the algebraic effect of putting nonzero values in the fourth row of the matrix, such as
(A,B,C,D). When you multiply the matrix by (Px, Py, Pz, 1) (or any multiple thereof) the fourth term in the
resulting point becomes APx + BPy + CPz + D, making it linearly dependent on each of the components of
P. After perspective division this term appears in the denominator of the point. Such a denominator is
exactly what is needed to produce the geometric effect of perspective projection onto a general plane, as
we show in the exercises.
The perspective transformation therefore carries a 3D point P into another 3D point P, according to:

( Px , Py , Pz ) N

Py aPz + b
Px
,N
,
Pz
Pz Pz

the perspective transformation

(7.11)

Where does the projection part come into play? Further along the pipeline the first two components of
this point are used for drawing: to locate in screen coordinates the position of the point to be drawn. The
third component is peeled off to be used for depth testing. As far as locating the point on the screen is
concerned, ignoring the third component is equivalent to replacing it by 0, as in:

N P , N P , aP + b N P , N P ,0
P P P P P
y

the projection

(7.12)

This is just what we did in Chapter 5 to project a point orthographically (meaning perpendicularly to
the viewplane) when setting up a camera for our first efforts at viewing a 3D scene. We will study
orthographic projections in full detail later. For now we can conclude:
(perspective projection) = (perspective transformation) + (orthographic projection)
This decomposition of a perspective projection into a specific transformation followed by a (trivial)
projection will prove very useful, both algorithmically and for understanding better what each point
actually experiences as it passes through the graphics pipeline. OpenGL does the transformation step
separately from the projection step. In fact it inserts clipping, perspective division, and one additional
mapping between them. We next look deeper into the transformation part of the process.
The Geometric Nature of the Perspective Transformation.
The perspective transformation alters 3D point P into another 3D point according to Equation 7.11, to
prepare it for projection. It is useful to think of it as causing a warping of 3D space, and to see how it
warps one shape into another. Very importantly, it preserves straightness and flatness, so lines
transform into lines, planes into planes, and polygonal faces into other polygonal faces. It also preserves
in-between-ness, so if point a is inside an object, the transformed point will also be inside the
transformed object. (Our choice of a suitable pseudodepth function was guided by the need to preserve
these properties.) The proof of these properties is developed in the exercises.
Of particular interest is how it transforms the cameras view volume, because if we are going to do
clipping in the warped space, we will be clipping against the warped view volume. The perspective
transformation shines in this regard: the warped view volume is a perfect shape for simple and efficient
clipping! Figure 7.28 suggests how the view volume and other shapes are transformed. The near plane W
at z = -N maps into the plane W at z = -1, and the far plane maps to the plane at z = +1. The top wall T is
tilted into the horizontal plane T so that is parallel to the z-axis. The bottom wall S becomes the

Chapter

3D viewing

November 23, 1999

page 20

horizontal S, and the two side walls become parallel to the z-axis. The cameras view volume is
transformed into a parallelepiped!

Figure 7.28. The view volume warped by the perspective transformation.

Its easy to prove how these planes map, because they all involve lines that are either parallel to the near
plane or that pass through the eye. Check these carefully:
Fact: Lines through the eye map into lines parallel to the z-axis. Proof: All points of such a line project
to a single point, say (x*, y*), on the viewplane. So all of the points along the line transform to all of the
points (x, y, z) with x = x*, y = y*, and z taking on all pseudodepth values between -1 and 1.
Fact: Lines perpendicular to the z-axis map to lines perpendicular to the z-axis. Proof: All points along
such a line have the same z-coordinate, so they all map to points with the same pseudodepth value.
Using these facts it is straightforward to derive the exact shape and dimensions of the warped view
volume.
The transformation also warps objects, like the blocks shown, into new shapes. Figure 7.28b shows a
block being projected onto the near plane. Suppose the top edge of its front face projects to y = 2 and the
top edge of its back face projects to y = 1. When this block is transformed, it becomes a truncated
pyramid: the top edge of its front face lies at y = 2, and the top edge of its back face lies at y = 1.Things
closer to the eye than the near plane become bigger, and things beyond the near plane become smaller.
The transformed object is smaller at the back than the front because the original object projects that way.
The x and y-coordinates of the transformed object are the x- and y-coordinates of the projection of the
original object. These are the coordinates you would encounter upon making an orthographic projection
of the transformed object. In a nutshell:
The perspective transformation warps objects so that, when viewed with an orthographic projection,
they appear the same as the original objects do when viewed with a perspective projection.
So all objects are warped into properly foreshortened shapes according to the rules of perspective
projection. Thereafter they can be viewed with an orthographic projection, and the correct picture is
produced.
We look more closely at the specific shape and dimensions of the transformed view volume.
Details of the Transformed View Volume, and mapping into the Canonical View Volume.
We want to put some numbers on the dimensions of the view volume before and after it is warped.
Consider the top plane, and suppose it passes through the point (left, top, -N) at z = -N as shown in Figure
7.29. Because it is composed of lines that pass through the eye and through points in the near plane all of
which have a y-coordinate of top, it must transform to the plane y = top. Similarly,

Figure 7.29. Details of the perspective transformation.

the bottom plane transforms to the y = bott plane;

the left plane transforms to the x = left plane;
the right plane transforms to the x = right.

We now know the transformed view volume precisely: a parallelepiped with dimensions that are related
to the cameras properties in a very simple way. This is a splendid shape to clip against as we shall see,
because its walls are parallel to the coordinate planes. But it would be an even better shape for clipping if
its dimensions didnt depend on the particular camera being used. OpenGL composes the perspective
transformation with another mapping that scales and shifts this parallelepiped into the canonical view
volume, a cube that extends from -1 to 1 in each dimension. Because this scales things differently in the

Chapter

3D viewing

November 23, 1999

page 21

x- and y- dimensions as it squashes the scene into a fixed volume it introduces some distortion, but the
distortion will be eliminated in the final viewport transformation.
The transformed view volume already extends from -1 to 1 in z, so it only needs to be scaled in the other
two dimensions. We therefore include a scaling and shift in both x and y to map the parallelepiped into
the canonical view volume. We first shift by -(right + left)/2 in x and by -(top + bott)/2 in y. Then we
scale by 2/(right - left) in x and by 2/(top - bott) in y. When the matrix multiplications are done (see the
exercises) we obtain the final matrix:

2N
right left
0
R=

00

right + left
right left
top + bott
top bott
( F + N )
FN
1

0
2N
top bott
0
0

0

2 FN
F N
0
0

(the projection matrix)

(7.13)

This is known as the projection matrix, and it performs the perspective transformation plus a scaling and
shifting to transform the cameras view volume into the canonical view volume. It is precisely the matrix
that OpenGL creates (and by which it multiplies the current matrix) when glFrustum(left,
right, bott, top, N, F) is executed. Recall that gluPerspective(viewAngle,
aspect, N, F) is usually used instead, as its parameters are more intuitive. This sets up the same
matrix, after computing values for top, bott, etc. using

top = N tan(

viewAngle / 2)
180

bott = -top, right = top * aspect, and left = -right.

Clipping Faces against the View Volume.
Recall from Figure 7.14 that clipping is performed after vertices have passed through the projection
matrix. It is done in this warped space because the canonical view volume is particularly well suited for
efficient clipping. Here we show how to exploit this, and we develop the details of the clipping
algorithm.
Clipping in the warped space works because a point lies inside the cameras view volume if and only if
its transformed version lies inside the canonical view volume. Figure 7.30a shows an example of clipping
in action. A triangle has vertices v1, v2, and v3. Vertex v3 lies outside the canonical view volume, CVV.
The clipper works on edges: it first clips edge v1v2, and finds that the entire edge lies inside CVV. Then it
clips edge v2v3, and records the new vertex a formed where the edge exits from the CVV. Finally it clips
edge v3v1 and records the new vertex where the edge enters the CVV. At the end of the process the
original triangle has become a quadrilateral with vertices v1v2a b. (We will see later that in addition to
identifying the locations of the new vertices, the pipeline also computes new color and texture parameters
at these new vertices.)
a). clip a triangle
b). clip an edge
Figure 7.30. Clipping against the canonical view volume.
The clipping problem is basically the problem of clipping a line segment against the CVV. We examined
such an algorithm, the Cyrus-Beck clipper, in Section 4.8.3. The clipper we develop here is similar to
that one, but of course it works in 3D rather than 2D.
Actually, it works in 4D. We will clip in the 4D homogeneous coordinate space called clip coordinates
in Figure 7.14. This is easier than it might seem, and it will nicely distinguish between points in front of,
and behind, the eye.
Suppose we want to clip the line segment AC shown in Figure 7.30b against the CVV. This means we are
given two points in homogeneous coordinates, A = (ax, ay, az, aw) and C = (cx, cy, cz, cw), and we want to

Chapter

3D viewing

November 23, 1999

page 22

determine which part of the segment lies inside the CVV. If the segment intersects the boundary of the
CVV we will need to compute the intersection point I = (Ix, Iy, Iz, Iw).
As with the Cyrus-Beck algorithm we view the CVV as six infinite planes, and consider where the given
edge lies relative to each plane in turn. We can represent the edge parametrically as A + (C-A)t. It lies at
A when t is 0, and at C when t is 1. For each wall of the CVV we first test whether A and C lie on the
same side of a wall: if they do there is no need to compute the intersection of the edge with the wall. If
they lie on opposite sides we locate the intersection point and clip off the part of the edge that lies
outside.
So we must be able to test whether a point is on the outside or inside of a plane. Take the plane x = -1,
for instance, which is one of the walls of the CVV. The point A lies to the right of it (on the inside) if

ax
> 1
aw

ax > aw

(a w + a x ) > 0 .

(7.14)

(When you multiply both sides of an inequality by a negative term you must reverse the direction of the
inequality. But we are ultimately dealing with only positive values of aw here - see the exercises.)
Similarly A is inside the plane x = 1 if

ax
>1
aw

( aw a x ) > 0

Blinn [blinn96] calls these quantities the boundary coordinates of point A, and he lists the six such
quantities that we work with as in Figure 7.31:
boundary coordinate homogeneous value clip plane
BC0
w+x
X=-1
BC1
w-x
X=1
w+y
Y=-1
BC2
BC3
w-y
Y=1
BC4
w+x
Z=-1
BC5
w-z
Z=1
Figure 7.31. The boundary codes computed for each end point of an edge.
We form these six quantities for A and again for C. If all six are positive the point lies inside the CVV. If
any are negative the point lies outside. If both points lie inside we have the same kind of trivial accept
we had in the Cohen Sutherland clipper of Section 3.3. If A and C lie outside on the same side
(corresponding BCs are negative) the edge must lie wholly outside the CVV.
Trivial accept: both endpoints lie inside the CVV (all 12 BCs are positive)
Trivial reject: both endpoints lie outside the same plane of the CVV.
If neither condition prevails we must clip segment AC against each plane individually. Just as with the
Cyrus-Beck clipper, we keep track of a candidate interval (CI) (see Figure 4.45), an interval of time
during which the edge might still be inside the CVV. Basically we know the converse: if t is outside the
CI we know for sure the edge is not inside the CVV. The CI extends from t = tin to tout.
We test the edge against each wall in turn. If the corresponding boundary codes have opposite signs we
know the edge hits the plane at some thit, which we then compute. If the edge is entering (is moving into
the inside of the plane at t increases) we update tin = max(old tin, thit), since it could not possibly be
entering at an earlier time than thit. Similarly, if the edge is exiting, we update tout = min(old tout, thit). If at
any time the CI is reduced to the empty interval (tout becomes > tin) we know the entire edge is clipped off
and we have an early out, which saves unnecessary computation.
It is straightforward to calculate the hit time of an edge with a plane. Write the edge parametrically in
homogeneous coordinates:

Chapter

3D viewing

November 23, 1999

page 23

edge(t) = (ax + (cx-aa) t, ay + (cy - ay) t, az + (cz - az) t, aw + (cw - aw) t)

If its the X = 1 plane, for instance, when the x-coordinate of A + (C-A)t is 1:

a x + (c x a x )t
=1
aw + (c w a w )t
This is easily solved for t, yielding

aw a x
(a w a x ) (c w c x )

(7.15)

Note that thit depends on only two boundary coordinates. Intersection with other planes yield similar
formulas.
This is easily put into code, as shown in Figure 7.32. This is basically the Liang Barsky algorithm
[liang84], with some refinements suggested by Blinn [blinn96]. The routine clipEdge(Point4 A,
Point4 C) takes two points in homogeneous coordinates (having fields x, y, z, and w) and returns 0 if
no part of AC lies in the CVV, and 1 otherwise. It also alters A and C so that when the routine is finished
A and C are the endpoints of the clipped edge.
The routine finds the six boundary coordinates for each endpoint and stores them in aBC[] and cBC[].
For efficiency it also builds an outcode for each point, which holds the signs of the six boundary codes
for that point. Bit i of As outcode holds a 0 if aBC[i] > 0 (A is inside the i-th wall) and a 1 otherwise.
Using these, a trivial accept occurs when both aOutcode and cOutcode are 0. A trivial reject occurs
when the bit-wise AND of the two outcodes is nonzero.
int clipEdge(Point4& A, Point4& C)
{
double tIn = 0.0, tOut = 1.0, tHit;
double aBC[6], cBC[6];
int aOutcode = 0, cOutcode = 0;
<.. find BCs for A and C ..>
<.. form outcodes for A and C ..>
if((aOutcode
return
if((aOutcode
return

& cOutcode) != 0) // trivial reject

0;
| cOutcode) == 0) // trivial accept
1;

for(int i = 0; i < 6; i++) // clip against each plane

{
if(cBC[i] < 0) // exits: C is outside
{
tHit = aBC[i]/(aBC[i] - cBC[i]);
tOut = MIN(tOut,tHit);
}
else if(aBC[i] < 0) //enters: A is outside
{
tHit = aBC[i]/(aBC[i] - cBC[i]);
tIn = MAX(tIn, tHit);
}
if(tIn > tOut) return 0; //CI is empty early out
}
// update the end points as necessary
Point4 tmp;
if(aOutcode != 0) // A is out: tIn has changed
{ // find updated A, (but dont change it yet)
tmp.x = A.x + tIn * (C.x - A.x);
tmp.y = A.y + tIn * (C.y - A.y);
tmp.z = A.z + tIn * (C.z - A.z);
tmp.w = A.w + tIn * (C.w - A.w);
Chapter

3D viewing

November 23, 1999

page 24

}
if(cOutcode != 0) // C is out: tOut has changed
{ // update C (using original value of A)
C.x = A.x + tOut * (C.x - A.x);
C.y = A.y + tOut * (C.y - A.y);
C.z = A.z + tOut * (C.z - A.z);
C.w = A.w + tOut * (C.w - A.w);
}
A = tmp; // now update A
return 1; // some of the edge lies inside the CVV
}
Figure 7.32. The edge clipper (as refined by Blinn).
In the loop that tests the edge against each plane, at most one of the BCs can be negative. (Why?) If A
has negative BC the edge must be entering at the hit point; if C has a negative BC the edge must be
exiting at the hit point. (Why?) (Blinn uses a slightly faster test by incorporating a mask that tests one bit
of an outcode.) Each time tIn or tOut are updated an early out is taken if tIn has become greater than
tOut.
When all planes have been tested, one or both of tIn and tOut have been altered (why?). A is updated
to A + (i - A) tIn if tIn has changed, and C is updated to A + (C - A) tOut if tOut has changed.
Blinn suggests pre-computing the BCs and outcode for every point to be processed. The eliminates the
need to re-compute these quantities when a vertex is an endpoint of more than one edge, as is often the
case.
Why did we clip against the canonical view volume?
Now that we have seen how easy it is to do clipping against the canonical view volume, we can see the
value of having transformed all objects of interest into it prior to clipping. There are two important
features of the CVV:
1.

It is parameter-free: the algorithm needs no extra information to describe the clipping volume. It uses
only the values -1 and 1. So the code itself can be highly tuned for maximum efficiency.

Its planes are aligned with the coordinate axes (after the perspective transformation is performed).
This means that we can determine which side of a plane a point lies on using a single coordinate, as
in ax > -1. If the planes were not aligned, an expensive dot product would be needed.

Why did we clip in homogeneous coordinates, rather than after the perspective division step?
This isnt completely necessary, but it makes the clipping algorithm clean, fast, and simple. Doing the
perspective divide step destroys information: if you have the values ax and aw explicitly you know of
course the signs of both of them. But given only the ratio ax/aw you can tell only whether ax and aw have
the same or opposite signs. Keeping values in homogeneous coordinates and clipping points closer to the
eye than the near plane automatically removes points that lie behind the eye, such as B in Figure 7.23.
Some perverse situations that necessitate clipping in homogeneous coordinates are described in
[blinn96, foley90]. They involve peculiar transformations of objects, or construction of certain surfaces,
where the original point (ax, ay, az, aw) has a negative fourth term, even though the point is in front of the
eye. None of the objects we discuss modeling here involve such cases. We conclude that clipping in
homogeneous coordinates, although usually not cricitcal, makes the algorithm fast and simple, and brings
it almost no cost.
Following the clipping operation perspective division is finally done, (as in Figure 7.14), and the 3-tuple
(x, y, z) is passed through the viewport transformation. As we discuss next, this transformation sizes and
shifts the x- and y- values so they are placed properly in the viewport, and makes minor adjustments on
the z- component (pseudodepth) to make it more suitable for depth testing.
The Viewport Transformation.

Chapter

3D viewing

November 23, 1999

page 25

As we have seen the perspective transformation squashes the scene into the canonical cube, as suggested
in Figure 7.33. If the aspect ratio of the cameras view volume (that is, the aspect ratio of the window on
the near plane) is 1.5, there is obvious distortion introduced when the perspective transformation scales
objects into a window with aspect ratio 1. But the viewport transformation can undo this distortion by
mapping a square into a viewport of aspect ratio 1.5. We normally set the aspect ratio of the viewport to
be the same as that of the view volume.

Figure 7.33. The viewport transformation restores aspect ratio.

We have encountered the OpenGL function glViewport(x, y, wid, ht) often before. It specifies
that the viewport will have lower left corner (x,y) in screen coordinates, and will be wid pixels wide
and ht pixels high. It thus specifies a viewport with aspect ratio wid/ht. The viewport transformation
also maps pseudodepth from the range -1 to 1 into the range 0 to 1.
Review Figure 7.14, which reveals the entire graphics pipeline. Each point P (which is usually one vertex
of a polygon) is passed through the following steps:
-

P is extended to a homogeneous 4-tuple by appending a 1;

This 4-tuple is multiplied by the modelview matrix, producing a 4-tuple giving the position in eye
coordinates;
The point is then multiplied by the projection matrix, producing a 4-tuple in clip coordinates;
The edge having this point as an endpoint is clipped;
Perspective division is performed, returning a 3-tuple;
The viewport transformation multiplies the 3-tuple by a matrix: the result (sx, sy, dz) is used for
drawing and depth calculations. (sx, sy) is the point in screen coordinates to be displayed; dz is a
measure of the depth of the original point from the eye of the camera.

Practice Exercises.
7.4.5. P projects where? Suppose the viewplane is given in camera coordinates by the equation Ax + By
+ Cz = D. Show that any point P projects onto this plane at the point given in homogeneous coordinates

P' = ( DPx , DPy , DPz , APx + BPy + CPz )

Do it using these steps.
a). Show that the projected point is the point at which the ray between the eye and P hits the given plane.
b). Show that the ray is given by Pt, and it hits the plane at t* = D/(APx + BPy + CPz).
c). Show that the projected point - the hit point - is therefore given properly above.
d). Show that for the near plane we use earlier, we obtain (x*, y*) as given by Equation 7.4.
7.4.6. A revealing approximate form for pseudodepth. Show that pseudodepth a + b/(-Pz), where a
and b are given in Equation 7.8, is well approximated by Equation 7.9 when N is much smaller than F.
7.4.7. Points at infinity in homogeneous coordinates. Consider the nature of the homogeneous
coordinate point (x, y, z, w) as w becomes smaller and smaller. For w = .01 it is (100x, 100y, 100z), for w
= 0.0001 it is (10000x, 10000y, 10000z), etc. It progresses out toward infinity in the direction (x, y, z).
The point with representation (x, y, z, 0) is in fact called a point at infinity. It is one of the advantages
of homogeneous coordinates that such an idealized point has a perfectly finite representation: in some
mathematical derivations this removes many awkward special cases. For instance, two lines will always
intersect, even if they are parallel [ayers67, semple52]. But other things dont work as well. What is the
difference of two points in homogeneous coordinates?
7.4.8. How does the perspective transformation affect lines and planes?
We must show that the perspective transformation preserves flatness and in-between-ness.
a). Argue why this is proven if we can show that a point P lying on the line between two points A and B
transforms to a point P that lies between the transformed versions of A and B.
b). Show that the perspective transformation does indeed produce a point P with the property just stated.
c). Show that each plane that passes through the eye maps to a plane that is parallel to the z-axis.
d). Show that each plane that is parallel to the z-axis maps to a plane parallel to the z- axis.
e). Show that relative depth is preserved;
7.4.9. The details of the transformed view volume. Show that the warped view volume has the
dimensions given in the five points above. You can use the facts developed in the preceding exercise.

Chapter

3D viewing

November 23, 1999

page 26

7.4.10. Show the final form of the projection matrix. The projection matrix is basically that of
Equation 7.10, followed by the shift and scaling described. If the matrix of Equation 7.10 is denoted as
M, and T represents the shifting matrix, and S the scaling matrix, show that the matrix product STM is
that given in Equation 7.13.
7.4.11. What Becomes of Points Behind the Eye? If the perspective transformation moves the eye off to
-infinity, what happens to points that lie behind the eye? Consider a line, P(t), that begins at a point in
front of the eye at t = 0 and moves to one behind the eye at t = 1.
a). Find its parametric form in homogeneous coordinates;
b). Find the parametric representation after it undergoes the perspective transformation;
c). Interpret it geometrically. Specifically state what the fourth homogeneous coordinate is geometrically.
A valuable discussion of this phenomenon is given in [blinn78].

Chapter

3D viewing

November 23, 1999

page 27

(For ECE660 - Fall, 1999)

Chapter 8 Rendering Faces for Visual Realism.

Goals of the Chapter
To add realism to drawings of 3D scenes.
To examine ways to determine how light reflects off of surfaces.
To render polygonal meshes that are bathed in light.
To see how to make a polygonal mesh object appear smooth.
To develop a simple hidden surface removal using a depth-buffer.
To develop methods for adding textures to the surfaces of objects.
To add shadows of objects to a scene.

Preview.
Section 8.1 motivates the need for enhancing the realism of pictures of 3D objects. Section 8.2
introduces various shading models used in computer graphics, and develops tools for computing the
ambient, diffuse, and specular light contributions to an objects color. It also describes how to set up
light sources in OpenGL, how to describe the material properties of surfaces, and how the OpenGL
graphics pipeline operates when rendering polygonal meshes.
Section 8.3 focuses on rendering objects modeled as polygon meshes. Flat shading, as well as
Gouraud and Phong shading, are described. Section 8.4 develops a simple hidden surface removal
technique based on a depth buffer. Proper hidden surface removal greatly improves the realism of
pictures.
Section 8.5 develops methods for painting texture onto the surface of an object, to make it appear
to be made of a real material such as brick or wood, or to wrap a label or picture of a friend around
it. Procedural textures, which create texture through a routine, are also described. The thorny issue
of proper interpolation of texture is developed in detail. Section 8.5.4 presents a complete program
that uses OpenGL to add texture to objects. The next sections discuss mapping texture onto curved
surfaces, bump mapping, and environment mapping, providing more tools for making a 3D scene
appear real.
Section 8.6 describes two techniques for adding shadows to pictures. The chapter finishes with a
number of Case Studies that delve deeper into some of these topics, and urges the reader to
experiment with them.

8.1. Introduction.
In previous chapters we have developed tools for modeling mesh objects, and for manipulating a
camera to view them and make pictures. Now we want to add tools to make these objects and others
look visually interesting, or realistic, or both. Some examples in Chapter 5 invoked a number of
OpenGL functions to produce shiny teapots and spheres apparently bathed in light, but none of the
underlying theory of how this is done was examined. Here we rectify this, and develop the lore of
rendering a picture of the objects of interest. This is the business of computing how each pixel of a
picture should look. Much of it is based on a shading model, which attempts to model how light
that emanates from light sources would interact with objects in a scene. Due to practical limitations
one usually doesnt try to simulate all of the physical principles of light scattering and reflection;
this is very complicated and would lead to very slow algorithms. But a number of approximate
models have been invented that do a good job and produce various levels of realism.
We start by describing a hierarchy of techniques that provide increasing levels of realism, in order
to show the basic issues involved. Then we examine how to incorporate each technique in an
application, and also how to use OpenGL to do much of the hard work for us.
At the bottom of the hierarchy, offering the lowest level of realism, is a wire-frame rendering.
Figure 8.1 shows a flurry of 540 cubes as wire-frames. Only the edges of each object are drawn, and
you can see right through an object. It can be difficult to see whats what. (A stereo view would
help a little.)

Chapter 8

November 30, 1999

page 1

Figure 8.1. A wire-frame rendering of a scene.

Figure 8.2 makes a significant improvement by not drawing any edges that lie behind a face. We
can call this a wire-frame with hidden surface removal rendering. Even though only edges are
drawn the objects now look solid and it is easy to tell where one stops and the next begins. Notice
that some edges simply end abruptly as they slip behind a face.

Figure 8.2. Wire-frame view with hidden surfaces removed.

(For the curious: This picture was made using OpenGL with its depth buffer enabled. For each mesh
object the faces were drawn in white using drawMesh(), and then the edges were drawn in black
using drawEdges(). Both routines were discussed in Chapter 6.)
The next step in the hierarchy produces pictures where objects appear to be in a scene,
illuminated by some light sources. Different parts of the object reflect different amounts of light
depending on the properties of the surfaces involved, and on the positions of the sources and the
cameras eye. This requires computing the brightness or color of each fragment rather than having
the user choose it. The computation requires the use of some shading model that determines the
proper amount of light that is reflected from each fragment.
Figure 8.3 shows a scene modeled with polygonal meshes: a buckyball rests atop two cylinders, and
the column rests on a floor. Part a shows the wire-frame version, and part b shows a shaded version
(with hidden surfaces removed). Those faces aimed toward the light source appear brighter than
those aimed away from the source. This picture shows flat shading: a calculation of how much
light is scattered from each face is computed at a single point, so all points on a face are rendered
with the same gray level.

Chapter 8

November 30, 1999

page 2

a).wire-frame

b). flat shaded

Figure 8.3. A mesh approximation shaded with a shading model. a). wire-frame view b). flat
shading,
The next step up is of course to use color. Plate 22 shows the same scene where the objects are
given different colors.
In Chapter 6 we discussed building a mesh approximation to a smoothly curved object. A picture of
such an object ought to reflect this smoothness, showing the smooth underlying surface rather
than the individual polygons. Figure 8.4 show the scene rendered using smooth shading. (Plate 23
shows the colored version.) Here different points of a face are drawn with different gray levels
found through an interpolation scheme known as Gouraud shading. The variation in gray levels is
much smoother, and the edges of polygons disappear, giving the impression of a smooth, rather than
a faceted, surface. We examine Gouraud shading in Section 8.3.

Figure 8.4. The scene rendered with smooth shading.

Highlights can be added to make objects look shiny. Figure 8.5 show the scene with specular light
components added. (Plate 24 shows the colored version.) The shinier an object is, the more
localized are its specular highlights, which often make an object appear to be made from plastic.

Figure 8.5. Adding specular highlights.

Another effect that improves the realism of a picture is shadowing. Figure 8.6 show the scene above
with shadows properly rendered, where one object casts a shadow onto a neighboring object. We
discuss how to do this in Section 8.6.

Figure 8.6. The scene rendered with shadows.

Adding texture to an object can produce a big step in realism. Figure 8.7 (and Plate 25) show the
scene with different textures painted on each surface. These textures can make the various
surfaces appear to made of some material such as wood, marble, or copper. And images can be
wrapped around an object like a decal.

Figure 8.7. Mapping textures onto surfaces.

There are additional techniques that improve realism. In Chapter 14 we study ray tracing in depth.
Although raytracing is a computationally expensive approach, it is easy to program, and produces
pictures that show proper shadows, mirror-like reflections, and the passage of light through
transparent objects.
In this chapter we describe a number of methods for rendering scenes with greater realism. We first
look at the classical lighting models used in computer graphics that make an object appear bathed in
light from some light sources, and see how to draw a polygonal mesh so that it appears to have a
smoothly curved surface. We then examine a particular hidden surface removal method - the one
that OpenGL uses - and see how it is incorporated into the rendering process. (Chapter 13 examines
a number of other hidden surface removal methods.) We then examine techniques for drawing
shadows that one object casts upon another, and for adding texture to each surface to make it appear
to be made of some particular material, or to have some image painted on it. We also examine
chrome mapping and environment mapping to see how to make a local scene appear to be
imbedded in a more global scene.

8.2. Introduction to Shading Models.

The mechanism of light reflection from an actual surface is very complicated, and it depends on
many factors. Some of these are geometric, such as the relative directions of the light source, the
Chapter 8

November 30, 1999

page 3

observer's eye, and the normal to the surface. Others are related to the characteristics of the surface,
such as its roughness, and color of the surface.
A shading model dictates how light is scattered or reflected from a surface. We shall examine some
simple shading models here, focusing on achromatic light. Achromatic light has brightness but no
color; it is only a shade of gray. Hence it is described by a single value: intensity. We shall see how
to calculate the intensity of the light reaching the eye of the camera from each portion of the object.
We then extend the ideas to include colored lights and colored objects. The computations are almost
identical to those for achromatic light, except that separate intensities of red, green, and blue
components are calculated.
A shading model frequently used in graphics supposes that two types of light sources illuminate the
objects in a scene: point light sources and ambient light. These light sources shine on the various
surfaces of the objects, and the incident light interacts with the surface in three different ways:

Some is absorbed by the surface and is converted to heat;

Some is reflected from the surface;
Some is transmitted into the interior of the object, as in the case of a piece of glass.

If all incident light is absorbed, the object appears black and is known as a black body. If all is
transmitted, the object is visible only through the effects of refraction, which we shall discuss in
Chapter 14.
Here we focus on the part of the light that is reflected or scattered from the surface. Some amount
of this reflected light travels in just the right direction to reach the eye, causing the object to be
seen. The fraction that travels to the eye is highly dependent on the geometry of the situation. We
assume that there are two types of reflection of incident light: diffuse scattering and specular
reflection.

Diffuse scattering, occurs when some of the incident light slightly penetrates the surface and is
re-radiated uniformly in all directions. Scattered light interacts strongly with the surface, and so
its color is usually affected by the nature of the surface material.
Specular reflections are more mirror-like and are highly directional: Incident light does not
penetrate the object but instead is reflected directly from its outer surface. This gives rise to
highlights and makes the surface look shiny. In the simplest model for specular light the
reflected light has the same color as the incident light. This tends to make the material look like
plastic. In a more complex model the color of the specular light varies over the highlight,
providing a better approximation to the shininess of metal surfaces. We discuss both models for
specular reflections.

Most surfaces produce some combination of the two types of reflection, depending on surface
characteristics such as roughness and type of material. We say that the total light reflected from the
surface in a certain direction is the sum of the diffuse component and the specular component. For
each surface point of interest we compute the size of each component that reaches the eye.
Algorithms are developed next that accomplish this.

8.2.1. Geometric Ingredients for Finding Reflected Light.

On the outside grows the furside, on the inside grows the skinside;
So the furside is the outside, and the skinside is the inside.
Herbert George Ponting, The Sleeping Bag
We need to find three vectors in order to compute the diffuse and specular components. Figure 8.8
shows the three principal vectors required to find the amount of light that reaches the eye from a
point P.

Figure 8.8. Important directions in computing reflected light.

1. The normal vector, m, to the surface at P.
Chapter 8

November 30, 1999

page 4

2. The vector v from P to the viewer's eye.

3. The vector s from P to the light source.
The angles between these three vectors form the basis for computing light intensities. These angles
are normally calculated using world coordinates, because some transformations (such as the
perspective transformation) do not preserve angles.
Each face of a mesh object has two sides. If the object is solid one is usually the inside and one is
the outside. The eye can then see only the outside (unless the eye is inside the object!), and it is
this side for which we must compute light contributions. But for some objects, such as the open box
of Figure 8.9, the eye might be able to see the inside of the lid. It depends on the angle between the
normal to that side, m2, and the vector to the eye, v. If the angle is less than 90o this side is visible.
Since the cosine of that angle is proportional to the dot product vm2, the eye can see this side only
if vm2 > 0.

Figure 8.9. Light computations are made for one side of each face.
We shall develop the shading model for a given side of a face. If that side of the face is turned
away from the eye there is normally no light contribution. In an actual application the rendering
algorithm must be told whether to compute light contributions from one side or both sides of a
given face. We shall see that OpenGL supports this.

8.2.2. Computing the Diffuse Component.

Suppose that light falls from a point source onto one side of a facet (a small piece of a surface). A fraction
of it is re-radiated diffusely in all directions from this side. Some fraction of the re-radiated part reaches
the eye, with intensity w denoted by Id. How does Id depend on the directions m, v, and s?
Because the scattering is uniform in all directions, the orientation of the facet, F, relative to the eye
is not significant. Therefore, Id is independent of the angle between m and v (unless v m < 0,
whereupon Id is zero.) On the other hand, the amount of light that illuminates the facet does depend
on the orientation of the facet relative to the point source: It is proportional to the area of the facet
that it sees, that is, the area subtended by a facet.
Figure 8.10a shows in cross section a point source illuminating a facet S when m is aligned with s.
In Figure 8.10b the facet is turned partially away from the light source through angle . The area
subtended is now only cos() as much as before, so that the brightness of S is reduced by this same
factor. This relationship between brightness and surface orientation is often called Lambert's law.
Notice that for near 0, brightness varies only slightly with angle, because the cosine changes
slowly there. As approaches 90, however, the brightness falls rapidly to 0.

Figure 8.10. The brightness depends on the area subtended.

Now we know that cos() is the dot product between normalized versions of s and m. We can
therefore adopt as the strength of the diffuse component:

I d = Is d

sm
| s|| m|

In this equation, Is is the intensity of the light source, and d is the diffuse reflection coefficient.
Note that if the facet is aimed away from the eye this dot product is negative and we want Id to
evaluate to 0. So a more precise computation of the diffuse component is:

s m , 0
| s|| m|

I d = I s d max

Chapter 8

(8.1)

November 30, 1999

page 5

This max term might be implemented in code (using the Vector3 methods dot() and length() see Appendix 3) by:
double tmp = s.dot(m); // form the dot product
double value = (tmp<0) ? 0 : tmp/(s.length() * m.length());
Figure 8.11 shows how a sphere appears when it reflects diffuse light, for six reflection coefficients:
0, 0.2, 0.4, 0.6, 0.8, and 1. In each case the source intensity is 1.0 and the background intensity is set
to 0.4. Note that the sphere is totally black when d is 0.0, and the shadow in its bottom half (where
the dot product above is negative) is also black.

Figure 8.11. Spheres with various reflection coefficients shaded with diffuse light.
(file: fig8.11.bmp )
In reality the mechanism behind diffuse reflection is much more complex than the simple model we
have adopted here. The reflection coefficient d depends on the wavelength (color) of the incident
light, the angle , and various physical properties of the surface. But for simplicity and to reduce
computation time, these effects are usually suppressed when rendering images. A reasonable
value for d is chosen for each surface, sometimes by trial and error according to the realism
observed in the resulting image.
In some shading models the effect of distance is also included, although it is somewhat
controversial. The light intensity falling on facet S in Figure 8.10 from the point source is known to
fall off as the inverse square of the distance between S and the source. But experiments have shown
that using this law yields pictures with exaggerated depth effects. (What is more, it is sometimes
convenient to model light sources as if they lie at infinity. Using an inverse square law in such a
case would quench the light entirely!) The problem is thought to be in the model: We model light
sources as point sources for simplicity, but most scenes are actually illuminated by additional
reflections from the surroundings, which are difficult to model. (These effects are lumped together
into an ambient light component.) It is not surprising, therefore, that strict adherence to a physical
law based on an unrealistic model can lead to unrealistic results.
The realism of most pictures is enhanced rather little by the introduction of a distance term. Some
approaches force the intensity to be inversely proportional to the distance between the eye and the
object, but this is not based on physical principles. It is interesting to experiment with such effects,
and OpenGL provides some control over this effect, as we see in Section 8.2.9, but we don't include
a distance term in the following development.

8.2.3. Specular Reflection.

Real objects do not scatter light uniformly in all directions, and so a specular component is added to
the shading model. Specular reflection causes highlights, which can add significantly to the realism
of a picture when objects are shiny. In this section we discuss a simple model for the behavior of
specular light due to Phong [Phong 1975]. It is easy to apply and OpenGL supports a good
approximation to it. Highlights generated by Phong specular light give an object a plastic-like
Chapter 8

November 30, 1999

page 6

appearance, so the Phong model is good when you intend the object to be made of shiny plastic or
glass. The Phong model is less successful with objects that are supposed to have a shiny metallic
surface, although you can roughly approximate them with OpenGL by careful choices of certain
color parameters, as we shall see. More advanced models of specular light have been developed that
do a better job of modeling shiny metals. These are not supported directly by OpenGLs rendering
process, so we defer a detailed discussion of them to Chapter 14 on ray tracing.
Figure 8.12a shows a situation where light from a source impinges on a surface and is reflected in
different directions. In the Phong model we discuss here, the amount of light reflected is greatest in
the direction of perfect mirror reflection, r, where the angle of incidence equals the angle of
reflection. This is the direction in which all light would travel if the surface were a perfect mirror.
At other near-by angles the amount of light reflected diminishes rapidly, as indicated by the relative
lengths of the reflected vectors. Part b shows this in terms of a beam pattern familiar in radar
circles. The distance from P to the beam envelope shows the relative strength of the light scattered
in that direction.
a).
b).
c).
Figure 8.12. Specular reflection from a shiny surface.
Part c shows how to quantify this beam pattern effect. We know from Chapter 5 that the direction r
of perfect reflection depends on both s and the normal vector m to the surface, according to:

r = s + 2

( s m)
m
| m|2

(the mirror-reflection direction)

(8.2)

For surfaces that are shiny but not true mirrors, the amount of light reflected falls off as the angle
between r and v increases. The actual amount of falloff is a complicated function of , but in the
Phong model it is said to vary as some power f of the cosine of , that is, according to (cos ())f, in
which f is chosen experimentally and usually lies between 1 and 200.
Figure 8.13 shows how this intensity function varies with for different values of f. As f increases,
the reflection becomes more mirror-like and is more highly concentrated along the direction r. A
perfect mirror could be modeled using f = , but pure reflections are usually handled in a different
manner, as described in Chapter 14.
similar to old15.14
Figure 8.13. Falloff of specular light with angle.
Using the equivalence of cos() and the dot product between r and v (after they are normalized), the
contribution Isp due to specular reflection is modeled by

I sp = I s s

r v
| r| | v|

(8.3)

where the new term s is the specular reflection coefficient. Like most other coefficients in the
shading model, it is usually determined experimentally. (As with the diffuse term, if the dot product
r v is found to be negative, Isp is set to zero.)
A boost in efficiency using the halfway vector. It can be expensive to compute the specular
term in Equation 8.3, since it requires first finding vector r and normalizing it. In practice an
alternate term, apparently first described by Blinn [blinn77], is used to speed up computation.
Instead of using the cosine of the angle between r and v, one finds a vector halfway between s and
v, that is, h = s + v, as suggested in Figure 8.14. If the normal to the surface were oriented along h
the viewer would see the brightest specular highlight. Therefore the angle between m and h can be
used to measure the falloff of specular intensity that the viewer sees. The angle is not the same as
(in fact is twice if the various vectors are coplanar - see the exercises), but this difference can be
compensated for by using a different value of the exponent f. (The specular term is not based on
physical principles anyway, so it is at least plausible that our adjustment to it yields acceptable
Chapter 8

November 30, 1999

page 7

results.) Thus it is common practice to base the specular term on cos() using the dot product of h
and m:

Figure 8.14. The halfway vector.

I sp = I s s

h m
max(0,
| h| | m|

{adjusted specular term}

(8.4)

Note that with this adjustment the reflection vector r need not be found, saving computation time.
In addition, if both the light source and viewer are very remote then s and v are constant over the
different faces of an object, so b need only be computed once.
Figure 8.15 shows a sphere reflecting different amounts of specular light. The reflection coefficient
s varies from top to bottom with values 0.25, 0.5, and 0.75, and the exponent f varies from left to
right with values 3, 6, 9, 25, and 200. (The ambient and diffuse reflection coefficients are 0.1 and
0.4 for all spheres.)

Figure 8.15. Specular reflection from a shiny surface.

The physical mechanism for specularly reflected light is actually much more complicated than the
Phong model suggests. A more realistic model makes the specular reflection coefficient dependent
on both the wavelength (i.e. the color) of the incident light and the angle of incidence i, (the
angle between vectors s and m in Figure 8.10) and couples it to a Fresnel term that describes the
physical characteristics of how light reflects off certain classes of surface materials. As mentioned,
OpenGL is not organized to include these effects, so we defer further discussion of them until
Chapter 14 on ray tracing, where we compute colors on a point by point basis, applying a shading
model directly.
Practice Exercises.
8.2.1. Drawing Beam Patterns. Draw beam patterns similar to that in Figure 8.12 for the cases f =
1, f = 10, and f = 100.
8.2.2. On the halfway vector. By examining the geometry displayed in Figure 8.14, show that =
2 if the vectors involved are coplanar. Show that this is not so if the vectors are non coplanar. See
also [fisher94]
8.2.3. A specular speed up. Schlick [schlick94] has suggested an alternative to the exponentiation
required when computing the specular term. Let D denote the dot product r v /| r|| v| in Equation
8.3. Schlick suggests replacing Df with f fDD + D , which is faster to compute. Plot these two functions
for values of D in [0,1] for various values of f and compare them. Pay particular attention to values
of D near 1, since this is where specular highlights are brightest.

Chapter 8

November 30, 1999

page 8

8.2.4. The Role of Ambient Light.

The diffuse and specular components of reflected light are found by simplifying the rules by
which physical light reflects from physical surfaces. The dependence of these components on the
relative positions of the eye, model, and light sources greatly improves the realism of a picture over
renderings that simply fill a wireframe with a shade.
But our desire for a simple reflection model leaves us with far from perfect renderings of a scene.
As an example, shadows are seen to be unrealistically deep and harsh. To soften these shadows, we
can add a third light component called ambient light.
With only diffuse and specular reflections, any parts of a surface that are shadowed from the point
source receive no light and so are drawn black! But this is not our everyday experience. The scenes
we observe around us always seem to be bathed in some soft non-directional light. This light arrives
by multiple reflections from various objects in the surroundings and from light sources that populate
the environment, such as light coming through a window, fluorescent lamps, and the like. But it
would be computationally very expensive to model this kind of light precisely.
Ambient Sources and Ambient Reflections
To overcome the problem of totally dark shadows, we imagine that a uniform background glow
called ambient light exists in the environment. This ambient light source is not situated at any
particular place, and it spreads in all directions uniformly. The source is assigned an intensity, Ia.
Each face in the model is assigned a value for its ambient reflection coefficient, a (often this is
the same as the diffuse reflection coefficient, d), and the term Iaa is simply added to whatever
diffuse and specular light is reaching the eye from each point P on that face. Ia and a are usually
arrived at experimentally, by trying various values and seeing what looks best. Too little ambient
light makes shadows appear too deep and harsh; too much makes the picture look washed out and
bland.
Figure 8.16 shows the effect of adding various amounts of ambient light to the diffuse light
reflected by a sphere. In each case both the diffuse and ambient sources have intensity 1.0, and the
diffuse reflection coefficient is 04. Moving from left to right the ambient reflection coefficient takes
on values 0.0, 0.1, 0.3, 0.5, and 0.7. With only a modest amount of ambient light the harsh shadows
on the underside of the sphere are softened and look more realistic. Too much ambient reflection,
on the other hand, suppresses the shadows excessively.

Figure 8.16. On the effect of ambient light.

8.2.5. Combining Light Contributions.

We can now sum the three light contributions - diffuse, specular, and ambient - to form the total
amount of light I that reaches the eye from point P:

I = I a a + I d d lambert + I sp s phong f

(8.5)

where we define the values

lambert = max(0,

sm
hm
), and phong = max(0,
)
| s|| m|
| h|| m|

(8.6)

I depends on the various source intensities and reflection coefficients, as well as on the relative
positions of the point P, the eye, and the point light source. Here we have given different names, Id
Chapter 8

November 30, 1999

page 9

and Isp, to the intensities of the diffuse and specular components of the light source, because
OpenGL allows you to set them individually, as we see later. In practice they usually have the same
value.
To gain some insight into the variation of I with the position of P, consider again Figure 8.10. I is
computed for different points P on the facet shown. The ambient component shows no variation
over the facet; m is the same for all P on the facet, but the directions of both s and v depend on P.
(For instance, s = S - P where S is the location of the light source. How does v depend on P and the
eye?) If the light source is fairly far away (the typical case), s will change only slightly as P
changes, so that the diffuse component will change only slightly for different points P. This is
especially true when s and m are nearly aligned, as the value of cos() changes slowly for small
angles. For remote light sources, the variation in the direction of the halfway vector h is also slight
as P varies. On the other hand, if the light source is close to the facet, there can be substantial
changes in s and h as P varies. Then the specular term can change significantly over the facet, and
the bright highlight can be confined to a small portion of the facet. This effect is increased when the
eye is also close to the facet -causing large changes in the direction of v - and when the exponent f
is very large.
Practice Exercise 8.2.4. Effect of the Eye Distance. Describe how much the various light
contributions change as P varies over a facet when a). the eye is far away from the facet and b).
when the eye is near the facet.

8.2.6. Adding Color.

It is straightforward to extend this shading model to the case of colored light reflecting from colored
surfaces. Again it is an approximation born from simplicity, but it offers reasonable results and is
serviceable.
Chapter 12 provides more detail and background on the nature of color, but as we have seen
previously colored light can be constructed by adding certain amounts of red, green, and blue light.
When dealing with colored sources and surfaces we calculate each color component individually,
and simply add them to form the final color of reflected light. So Equation 8.5 is applied three
times:

Ir = Ia r ar + Idr dr lambert + Ispr s r phong f

Ig = Iag ag + Idg dg lambert + Ispg sg phong f

(8.7)

Ib = Ia b ab + Idb db lambert + Ispb sb phong f

(where lambert and phong are given in Equation 8.6) to compute the red, green, and blue
components of reflected light. Note that we say the light sources have three types of color:
ambient = (Iar, Iag, Iab), diffuse = (Idr, Idg, Idb), and specular = (Ispr, Ispg, Ispb ). Usually the diffuse and
specular light colors are the same. Note also that the lambert and phong terms do not depend on
which color component is being computed, so they need only be computed once. To pursue this
approach we need to define nine reflection coefficients:
ambient reflection coefficients:
diffuse reflection coefficients: ,
specular reflection coefficients: ,

ar, ag, and ab,

dr, dg, and db
sr, sg, and sb

The ambient and diffuse reflection coefficients are based on the color of the surface itself. By
color of a surface we mean the color that is reflected from it when the illumination is white light:
a surface is red if it appears red when bathed in white light. If bathed in some other color it can
exhibit an entirely different color. The following examples illustrate this.
Example 8.2.1. The color of an object. If we say that the color of a sphere is 30% red, 45% green,
and 25% blue, it makes sense to set its ambient and diffuse reflection coefficients to (0.3K, 0.45K,
0.25K), where K is some scaling value that determines the overall fraction of incident light that is
reflected from the sphere. Now if it is bathed in white light having equal amounts of red, green, and
blue (Isr = Isg = Isb = I) the individual diffuse components have intensities Ir = 0.3 K I, Ig = 0.45 K I, Ib
= 0.25 K I, so as expected we see a color that is 30% red, 45% green, and 25% blue.
Chapter 8

November 30, 1999

page 10

Example 8.2.2. A reddish object bathed in greenish light. Suppose a sphere has ambient and
diffuse reflection coefficients (0.8 , 0.2, 0.1 ), so it appears mostly red when bathed in white light.
We illuminate it with a greenish light Is = (0.15, 0.7, 0.15). The reflected light is then given by
(0.12, 0.14, 0.015), which is a fairly even mix of red and green, and would appear yellowish (as we
discuss further in Chapter 12).
The color of specular light. Because specular light is mirror-like, the color of the specular
component is often the same as that of the light source. For instance, it is a matter of experience
that the specular highlight seen on a glossy red apple when illuminated by a yellow light is yellow
rather than red. This is also observed for shiny objects made of plastic-like material. To create
specular highlights for a plastic surface the specular reflection coefficients, sr, sg, and sb used in
Equation 8.7 are set to the same value, say s, so that the reflection coefficients are gray in nature
and do not alter the color of the incident light. The designer might choose s = 0.5 for a slightly
shiny plastic surface, or s = 0.9 for a highly shiny surface.
Objects made of different materials.
A careful selection of reflection coefficients can make an object appear to be made of a specific
material such as copper, gold, or pewter, at least approximately. McReynolds and Blythe
[mcReynolds97] have suggested using the reflection coefficients given in Figure 8.17. Plate ???
shows several spheres modelled using these coefficients. The spheres do appear to be made of
different materials. Note that the specular reflection coefficients have different red, green, and blue
components, so the color of specular light is not simply that of the incident light. But McReynolds
and Blythe caution users that, because OpenGLs shading algorithm incorporates a Phong specular
component, the visual effects are not completely realistic. We shall revisit the issue in Chapter 14
and describe the more realistic Cook-Torrance shading approach..
Material
exponent: f
ambient: ar, ag, ab diffuse: dr, dg,db
specular: sr, sg,sb
Black
Plastic

0.0
0.01
0.50
0.0
0.01
0.50
0.0
0.01
0.50
Brass
0.329412
0.780392
0.992157
0.223529
0.568627
0.941176
0.027451
0.113725
0.807843
Bronze
0.2125
0.714
0.393548
0.1275
0.4284
0.271906
0.054
0.18144
0.166721
Chrome
0.25
0.4
0.774597
0.25
0.4
0.774597
0.25
0.4
0.774597
Copper
0.19125
0.7038
0.256777
0.0735
0.27048
0.137622
0.0225
0.0828
0.086014
Gold
0.24725
0.75164
0.628281
0.1995
0.60648
0.555802
0.0745
0.22648
0.366065
Pewter
0.10588
0.427451
0.3333
0.058824
0.470588
0.3333
0.113725
0.541176
0.521569
Silver
0.19225
0.50754
0.508273
0.19225
0.50754
0.508273
0.19225
0.50754
0.508273
Polished
0.23125
0.2775
0.773911
Silver
0.23125
0.2775
0.773911
0.23125
0.2775
0.773911
Figure 8.17. Parameters for common materials [mcReynolds97].

Chapter 8

November 30, 1999

27.8974

25.6

76.8

12.8

51.2

9.84615

51.2

89.6

page 11

Plate 26. Shiny spheres made of different materials.

8.2.7. Shading and the Graphics pipeline.

At which step in the graphics pipeline is shading performed? And how is it done? Figure 8.18
shows the pipeline again. The key idea is that the vertices of a mesh are sent down the pipeline
along with their associated vertex normals, and all shading calculations are done on vertices.
(Recall that the draw() method in the Mesh class sends a vertex normal along with each vertex, as
in Figure 6.15.)

Figure 8.18. The graphics pipeline revisited.

The figure shows a triangle with vertices v0, v1, and v2 being rendered. Vertex vi has the normal
vector mi associated with it. These quantities are sent down the pipeline with calls such as:
glBegin(GL_POLYGON);
for(int i = 0; i < 3; i++)
{
glNormal3f(norm[i].x, norm[i].y, norm[i].z);
glVertex3f(pt[i].x, pt[i].y, pt[i].z);
}
glEnd();
The Call to glNormal3f() sets the current normal vector, which is applied to all vertices
subsequently sent using glVertex3f(). It remains current until changed with another call to
glNormal3f(). For this code example a new normal is associated with each vertex.
The vertices are transformed by the modelview matrix, M, effectively expressing them in camera
(eye) coordinates. The normal vectors are also transformed, but vectors transform differently from
points. As shown in Section 6.5.3, transforming points of a surface by a matrix M causes the normal
-T
-T
m at any point to become the normal M m on the transformed surface, where M is the transpose of
the inverse of M. OpenGL automatically performs this calculation on normal vectors.
As we discuss in the next section OpenGL allows you to specify various light sources and their
locations. Lights are objects too, and the light source positions are also transformed by the
modelview matrix.
So all quantities end up after the modelview transformation being expressed in camera coordinates.
At this point the shading model of Equation 8.7 is applied, and a color is attached to each vertex.
The computation of this color requires knowledge of vectors m, s, and v, but these are all available
at this point in the pipeline. (Convince yourself of this).
Progressing farther down the pipeline, the pseudodepth term is created and the vertices are passed
through the perspective transformation. The color information tags along with each vertex. The
clipping step is performed in homogeneous coordinates as described earlier. This may alter some of
the vertices. Figure 8.19 shows the case where vertex v1 of the triangle is clipped off, and two new
vertices, a and b, are created. The triangle becomes a quadrilateral. The color at each of the new
vertices must be computed, since it is needed in the actual rendering step.

Figure 8.19. Clipping a polygon against the (warped) view volume.

Chapter 8

November 30, 1999

page 12

The color at each new vertex is usually found by interpolation. For instance, suppose that the color
at v0 is (r0, g0, b0) and the color at v1 is (r1, g1, b1). If the point a is 40% of the way from v0 to v1 the
color associated with a is a blend of 60% of (r0, g0, b0) and 40% of (r1, g1, b1). This is expressed as
color at point a = (lerp(r0, r1, 0.4), lerp(g0, g1, 0.4), lerp(b0, b1, 0.4))

(8.8)

where we use the convenient function lerp() (short for linear interpolation - recall tweening in
Section 4.5.3) defined by:
lerp(G, H, f) = G + (H - G)f

(8.9)

Its value lies at fraction f of the way from G to H.1

The vertices are finally passed through the viewport transformation where they are mapped into
screen coordinates (along with pseudodepth which now varies between 0 and 1). The quadrilateral
is then rendered (with hidden surface removal), as suggested in Figure 8.19. We shall say much
more about the actual rendering step.

8.2.8. Using Light Sources in OpenGL.

OpenGL provides a number of functions for setting up and using light sources, as well as for
specifying the surface properties of materials. It can be daunting to absorb all of the many possible
variations and details, so we describe the basics here. In this section we discuss how to establish
different kinds of light sources in a scene. In the next section we look at ways to characterize the
reflective properties of the surfaces of an object.
Creating a light source.
OpenGL allows you to define up to eight sources, which are referred to through names
GL_LIGHT0, GL_LIGHT1, etc. Each source is invested with various properties, and must be
enabled. Each property has a default value. For example, to create a source located at (3, 6, 5) in
world coordinates, use2:
GLfloat myLightPosition[] = {3.0, 6.0, 5.0, 1.0};
glLightfv(GL_LIGHT0, GL_POSITION, myLightPosition);
glEnable(GL_LIGHTING); // enable
glEnable(GL_LIGHT0); // enable this particular source
The array myLightPosition[] (use any name you wish for this array) specifies the location of
the light source, and is passed to glLightfv() along with the name GL_LIGHT0 to attach it to the
particular source GL_LIGHT0.
Some sources, such as a desk lamp, are in the scene, whereas others, like the sun, are infinitely
remote. OpenGL allows you to create both types by using homogeneous coordinates to specify light
position:
(x, y, z, 1) : a local light source at the position (x, y, z)
(x, y, z, 0): a vector to an infinitely remote light source in the direction (x, y, z)
Figure 8.20 shows a local source positioned at (0, 3, 3, 1) and a remote source located along
vector (3, 3, 0, 0). Infinitely remote light sources are often called directional. There are
computational advantages to using directional light sources, since the direction s in the calculations
of diffuse and specular reflections is constant for all vertices in the scene. But directional light
sources are not always the correct choice: some visual effects are properly achieved only when a
light source is close to an object.

1 In Section 8.5 we discuss replacing linear interpolation by hyperbolic interpolation as a more accurate way
to form the colors at the new vertices formed by clipping.
2 Here and elsewhere the type float would most likely serve as well as GLfloat. But using GLfloat
makes your code more portable to other OpenGL environments.

Chapter 8

November 30, 1999

page 13

Figure 8.20. A local source and an infinitely remote source.

You can also spell out different colors for a light source. OpenGL allows you to assign a different
color to three types of light that a source emits: ambient, diffuse, and specular. It may seem
strange to say that a source emits ambient light. It is still treated as in Equation 8.7: a global omnidirectional light that bathes the entire scene. The advantage of attaching it to a light source is that it
can be turned on and off as an application proceeds. (OpenGL also offers a truly ambient light, not
associated with any source, as we discuss later in connection with lighting models.)
Arrays are defined to hold the colors emitted by light sources, and they are passed to
glLightfv().
GLfloat amb0[] = {0.2, 0.4, 0.6, 1.0}; // define some colors
GLfloat diff0[] = {0.8, 0.9, 0.5, 1.0};
GLfloat spec0[] = { 1.0, 0.8, 1.0, 1.0};
glLightfv(GL_LIGHT0, GL_AMBIENT, amb0); // attach them to LIGHT0
glLightfv(GL_LIGHT0, GL_DIFFUSE, diff0);
glLightfv(GL_LIGHT0, GL_SPECULAR, spec0);
Colors are specified in so-called RGBA format, meaning red, green, blue, and alpha. The alpha
value is sometimes used for blending two colors on the screen. We discuss it in Chapter 10. For our
purposes here it is normally 1.0.
Light sources have various default values. For all sources:
default ambient = (0, 0, 0, 1);

dimmest possible: black

For light source LIGHT0:

default diffuse = (1, 1, 1, 1);
default specular = (1, 1, 1, 1);

brightest possible: white

whereas for the other sources the diffuse and specular values have defaults of black.
Spotlights.
Light sources are point sources by default, meaning that they emit light uniformly in all directions.
But OpenGL allows you to make them into spotlights, so they emit light in a restricted set of
directions. Figure 8.21 shows a spotlight aimed in direction d, with a cutoff angle of .
No light is seen at points lying outside the cutoff cone. For vertices such as P that lie inside the

cone, the amount of light reaching P is attenuated by the factor cos ( ) where is the angle
between d and a line from the source to P, and is an exponent chosen by the user to give the
desired fall-off of light with angle.

Figure 8.21. Properties of a spotlight.

The parameters for a spotlight are set using glLightf() to set a single value, and glLightfv()
to set a vector:
o

glLightf(GL_LIGHT0, GL_SPOT_CUTOFF, 45.0); // a cutoff angle of 45

glLightf(GL_LIGHT0,GL_SPOT_EXPONENT, 4.0); // = 4.0
GLfloat dir[] = {2.0, 1.0, -4.0}; // the spotlights direction
glLightfv(GL_LIGHT0,GL_SPOT_DIRECTION, dir);

The default values for these parameters are d = (0,0,-1), = 180 , and = 0, which makes a source
an omni-directional point source.
o

Attenuation of light with distance.

OpenGL also allows you to specify how rapidly light diminishes with distance from a source.
Although we have downplayed the importance of this dependence, it can be interesting to

Chapter 8

November 30, 1999

page 14

experiment with different fall-off rates, and to fine tune a picture. OpenGL attenuates the strength
of a positional3 light source by the following attenuation factor:

atten =

1
k c + kl D + k q D 2

(8.11)

where kc, kl, and kq are coefficients and D is the distance between the lights position and the vertex
in question. This expression is rich enough to allow you to model any combination of constant,
linear, and quadratic (inverse square law) dependence on distance from a source. These parameters
are controlled by function calls:
glLightf(GL_LIGHT0, GL_CONSTANT_ATTENUATION, 2.0);
and similarly for GL_LINEAR_ATTENUATION, and GL_QUADRATIC_ATTENUATION. The
default values are kc = 1, kl = 0, and kq = 0, which eliminate any attenuation.
Lighting model.
OpenGL allows three parameters to be set that specify general rules for applying the shading model.
These parameters are passed to variations of the function glLightModel.
a). The color of global ambient light. You can establish a global ambient light source in a scene
that is independent of any particular source. To create this light, specify its color using:
GLfloat amb[] = {0.2, 0.3, 0.1, 1.0};
glLightModelfv(GL_LIGHT_MODEL_AMBIENT, amb);
This sets the ambient source to the color (0.2, 0.3,0.1). The default value is (0.2, 0.2, 0.2, 1.0), so
this ambient light is always present unless you purposely alter it. This makes objects in a scene
visible even if you have not invoked any of the lighting functions.
b). Is the viewpoint local or remote? OpenGL computes specular reflections using the halfway
vector h = s + v described in Section 8.2.3. The true directions s and v are normally different at
each vertex in a mesh (visualize this). If the light source is directional then s is constant, but v still
varies from vertex to vertex. Rendering speed is increased if v is made constant for all vertices. This
is the default: OpenGL uses v = (0, 0, 1), which points along the positive z-axis in camera
coordinates. You can force the pipeline to compute the true value of v for each vertex by executing:
glLightModeli(GL_LIGHT_MODEL_LOCAL_VIEWER, GL_TRUE);
c). Are both sides of a polygon shaded properly? Each polygonal face in a model has two sides.
When modeling we tend to think of them as the inside and outside surfaces. The convention is
to list the vertices of a face in counter-clockwise (CCW) order as seen from outside the object. Most
mesh objects represent solids that enclose space, so there is a well defined inside and outside. For
such objects the camera can only see the outside surface of each face (assuming the camera is not
inside the object!). With proper hidden surface removal the inside surface of each face is hidden
from the eye by some closer face.
OpenGL has no notion of inside and outside. It can only distinguish between front faces and
back faces. A face is a front face if its vertices are listed in counter-clockwise (CCW) order as
seen by the eye4. Figure 8.22a shows the eye viewing a cube, which we presume was modeled using
the CCW ordering convention. Arrows indicate the order in which the vertices of each face are
passed to OpenGL (in a glBegin(GL_POLYGON);...; glEnd() block). For a space-enclosing
object all faces that are visible to the eye are therefore front faces, and OpenGL draws them

This attenuation factor is disabled for directional light sources, since they are infinitely remote.
You can reverse this sense with glFrontFace(GL_CW), which decrees that a face is a front face only if its
vertices are listed in clock-wise order. The default is glFrontFace(GL_CCW).
4

Chapter 8

November 30, 1999

page 15

properly with the correct shading. OpenGL also draws the back faces5, but they are ultimately
hidden by closer front faces.
a).
b).
Figure 8.22. OpenGLs definition of a front face.
Things are different in part b, which shows a box with a face removed. Again arrows indicate the
order in which vertices of a face are sent down the pipeline. Now three of the visible faces are back
faces. By default OpenGL does not shade these properly. To coerce OpenGL to do proper shading
of back faces, use:
glLightModeli(GL_LIGHT_MODEL_TWO_SIDE, GL_TRUE);
Then OpenGL reverses the normal vectors of any back-face so that they point toward the viewer,
and it performs shading computations properly. Replace GL_TRUE with GL_FALSE (the default) to
turn off this facility.
Note: Faces drawn by OpenGL do not cast shadows, so the back faces receive the same light from a
source even though there may be some other face between them and the source.
Moving light sources.
Recall that light sources pass through the modelview matrix just as vertices do. Therefore lights can
be repositioned by suitable uses of glRotated() and glTranslated(). The array position
specified using glLightfv(GL_LIGHT0, GL_POSITION, position)is modified by the
modelview matrix in effect at the time glLightfv() is called. So to modify the light position with
transformations, and independently move the camera, imbed the light positioning command in a
push/pop pair, as in:
void display()
{
GLfloat position[] = {2, 1, 3, 1}; //initial light position
<.. clear color and depth buffers ..>
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
glPushMatrix();
glRotated(...); // move the light
glTranslated(...);
glLightfv(GL_LIGHT0, GL_POSITION, position);
glPopMatrix();
gluLookAt(...); // set the camera position
<.. draw the object ..>
glutSwapBuffers();
}
On the other hand, to have the light move with the camera, use:
GLfloat pos[] = {0,0,0,1};
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
glLightfv(GL_LIGHT0, GL_POSITION, position);// light at (0,0,0)
gluLookAt(...); // move the light and the camera
<.. draw the object ..>
This establishes the light to be positioned at the eye (like a minors lamp), and the light moves with
the camera.

8.2.9. Working with Material Properties in OpenGL.

5 You can improve performance by instructing OpenGL to skip rendering of back faces, with
glCullFace(GL_BACK); glEnable(GL_CULL_FACE);

Chapter 8

November 30, 1999

page 16

You can see the effect of a light source only when light reflects off an objects surface. OpenGL
provides ways to specify the various reflection coefficients that appear in Equation 8.7. They are set
with variations of the function glMaterial, and they can be specified individually for front faces
and back faces (see the discussion concerning Figure 8.22). For instance,
GLfloat myDiffuse[] = {0.8, 0.2, 0.0, 1.0};
glMaterialfv(GL_FRONT, GL_DIFFUSE, myDiffuse);
sets the diffuse reflection coefficient (dr, dg, db) = (0.8, 0.2, 0.0) for all subsequently specified
front faces. Reflection coefficients are specified as a 4-tuple in RBGA format, just like a color. The
first parameter of glMaterialfv() can take on values:
GL_FRONT : set the reflection coefficient for front faces
GL_BACK: set it for back faces
GL_FRONT_AND_BACK: set it for both front and back faces
The second parameter can take on values:
GL_AMBIENT: set the ambient reflection coefficients
GL_DIFFUSE: set the diffuse reflection coefficients
GL_SPECULAR: set the specular reflection coefficients
GL_AMBIENT_AND_DIFFUSE: set both the ambient and diffuse reflection coefficients to the same
values. This is for convenience, since the ambient and diffuse coefficients are so often chosen to be
the same.
GL_EMISSION: set the emissive color of the surface.
The last choice sets the emissive color of a face, causing it to glow in the specified color,
independent of any light source.
Putting it all together.
We now extend Equation 8.7 to include the additional contributions that OpenGL actually
calculates. The total red component is given by:

I r = er + I mr ar + atteni spot i I i a r ar + I i dr dr lambert i + I i spr s r phongi f

(8.12)
Expressions for the green and blue components are similar. The emissive light is er , and Imr is the
global ambient light introduced in the lighting model. The summation denotes that the ambient,
diffuse, and specular contributions of all light sources are summed. For the i-th source atteni is the
attenuation factor as in Equation 8.10, spoti is the spotlight factor (see Figure 8.21), and lamberti
and phongi are the familiar diffuse and specular dot products. All of these terms must be
recalculated for each source.
Note: If Ir turns out to have a value larger than 1.0, OpenGL clamps it to 1.0: the brightest any light
component can be is 1.0.

8.2.10. Shading of Scenes Specified by SDL.

The scene description language SDL introduced in Chapter 5 supports the loading of material
properties into objects, so that they can be shaded properly. For instance,
light 3 4 5 .8 .8 .8 ! bright white light at (3, 4, 5)
background 1 1 1 ! white background
globalAmbient .2 .2 .2 ! a dark gray global ambient light
ambient .2 .6 0
diffuse .8 .2. 1 ! red material
specular 1 1 1 ! bright specular spots the color of the source
exponent 20 !set the Phong exponent
scale 4 4 4 sphere
describes a scene containing a sphere with material properties (see Equation 8.7):

Chapter 8

November 30, 1999

page 17

ambient reflection coefficients:

diffuse reflection coefficients:
specular reflection coefficients:
and Phong exponent f = 20.

(ar, ag, ab) = (.2, 0.6, 0),

(dr, dg, db) = (0.8, 0.2, 1.0),
(sr, sg, sb) = (1.0, 1.0, 1.0)

The light source is given a color of (0.8, 0.8, 0.8) for both its diffuse and specular components.
There is a global ambient term (Iar, Iag, Iab) = (0.2, 0.2, 0.2).
The current material properties are loaded into each objects mtrl field at the time it is created (see
the end of Scene :: getObject() in Shape.cpp of Appendix 4). When an object draws itself
using its drawOpenGL() method, it first passes its material properties to OpenGL (see Shape::
tellMaterialsGL()), so that at the moment it is actually drawn OpenGL has these properties in
its current state.
In Chapter 14 when raytracing we shall use each objects material field in a similar way to acquire
the material properties and do proper shading.

8.3. Flat Shading and Smooth Shading.

Different objects require different shading effects. In Chapter 6 we modeled a variety of shapes
using polygonal meshes. For some, like the barn or buckyball, we want to see the individual faces in
a picture, but for others, like the sphere or chess pawn, we want to see the underlying surface that
the faces approximate.
In the modeling process we attached a normal vector to each vertex of each face. If a certain face is
to appear as a distinct polygon we attach the same normal vector to all of its vertices; the normal
vector chosen is the normal direction to the plane of the face. On the other hand, if the face is
supposed to approximate an underlying surface we attach to each vertex the normal to the
underlying surface at that point.
We examine now how the normal vector information at each vertex is used to perform different
kinds of shading. The main distinction is between a shading method that accentuates the individual
polygons (flat shading) and a method that blends the faces to de-emphasize the edges between them
(smooth shading). There are two kinds of smooth shading, called Gouraud and Phong shading, and
we shall discuss both.
For both kinds of shading the vertices are passed down the graphics pipeline, shading calculations
are performed to attach a color to each vertex, and ultimately the vertices of the face are converted
to screen coordinates and the face is painted pixel by pixel with the appropriate color.
Painting a Face.
The face is colored using a polygon-fill routine. Filling a polygon is very simple, although fine
tuning the fill algorithm for highest efficiency can get complex. (See Chapter 10.) Here we look at
the basics, focusing on how the color of each pixel is set.
A polygon-fill routine is sometimes called a tiler, because it moves over the polygon pixel by pixel,
coloring each pixel as appropriate, as one would lay down tiles on a parquet floor. Specifically, the
pixels in a polygon are visited in a regular order, usually scan-line by scan-line from the bottom to
the top of the polygon, and across each scan-line from left to right.
We assume here that the polygons of interest are convex. A tiler designed to fill only convex
polygons can be made highly efficient, since at each scan-line there is a single unbroken run of
pixels that lie inside the polygon. Most implementations of OpenGL exploit this and always fill
convex polygons correctly, but do not guarantee to fill non-convex polygons properly. See the
exercises for more thoughts on convexity.
Figure 8.23 shows an example where the face is a convex quadrilateral. The screen coordinates of
each vertex are noted. The lowest and highest points on the face are ybott and ytop, respectively. The
tiler first fills in the row at y = ybott (in this case a single pixel), then the one at ybott+ 1, etc. At each
scan-line, say ys in the figure, there is a leftmost pixel, xleft, and a rightmost, xright. The tiler moves
from xleft to xright, placing the desired color in each pixel. So the tiler is implemented as a simple
double loop:
Chapter 8

November 30, 1999

page 18

Figure 8.23. Filling a polygonal face with color.

// for each scan-line
for (int y = ybott; y <= ytop; y++)
{
<.. find xleft and xright ..>
for (int x = xleft; x <= xright; x++)// fill across the scan-line
{
<.. find the color c for this pixel ..>
<.. put c into the pixel at (x, y) ..>
}
}
(We shall see later how hidden surface removal is easily accomplished within this double loop as
well.) The principal difference between flat and smooth shading is the manner in which the color c
is determined at each pixel.

8.3.1. Flat Shading.

When a face is flat (like the roof of a barn) and the light sources are quite distant the diffuse light
component varies little over different points on the roof (the lambert term in Equation 8.6 is nearly
the same at each vertex of the face). In such cases it is reasonable to use the same color for every
pixel covered by the face. OpenGL offers a rendering mode in which the entire face is drawn with
the same color. Although a color is passed down the pipeline as part of each vertex of the face, the
painting algorithm uses only one of them (usually that of the first vertex in the face). So the
command above, <find the color c for this pixel>, is not inside the loops but
instead appears just prior to the loops , setting c to the color of one of the vertices. (Using the same
color for every pixel tends to make flat shading quite fast.)
Flat shading is established in OpenGL using:
glShadeModel(GL_FLAT);
Figure 8.24 shows a buckyball and a sphere rendered using flat shading. The individual faces are
clearly visible on both objects. The sphere is modeled as a smooth object, but no smoothing is
taking place in the rendering, since the color of an entire face is set to that of only one vertex.

Figure 8.24. Two meshes rendered using flat shading.

Edges between faces actually appear more pronounced than they are, due to a phenomenon in the
eye known as lateral inhibition, first described by Ernst Mach6. When there is a discontinuity in
intensity across an object the eye manufactures a Mach band at the discontinuity, and a vivid edge

Ernst Mach (1838-1916), an Austrian physicist, whose early work strongly influenced the theory of
relativity.
Chapter 8

November 30, 1999

page 19

is seen (as discussed further in the exercises). This exaggerates the polygonal look of mesh
objects rendered with flat shading.
Specular highlights are rendered poorly with flat shading, again because an entire face is filled with
a color that was computed at only one vertex. If there happens to be a large specular component at
the representative vertex, that brightness is drawn uniformly over the entire face. If a specular
highlight doesnt fall on the representative point, it is missed entirely. For this reason, there is little
incentive for including the specular reflection component in the shading computation.

8.3.2.Smooth Shading.
Smooth shading attempts to de-emphasize edges between faces by computing colors at more points
on each face. There are two principal types of smooth shading, called Gouraud shading and Phong
shading [gouraud71, phong75]. OpenGL does only Gouraud shading, but we describe both of them.
Gouraud shading computes a different value of c for each pixel. For the scanline at ys (in figure
8.23) it finds the color at the leftmost pixel, colorleft by linear interpolation of the colors at the top
and bottom of the left edge7. For the scan-line at ys the color at the top is color4 and that at the
bottom is color1, so colorleft would be calculated as (recall Equation 8.9):
colorleft = lerp(color1, color4, f)

(8.13)

where fraction f , given by

f =

ys ybott
y4 ybott

varies between 0 and 1 as ys varies from ybott to y4. Note that Equation 8.13 involves three
calculations since each color quantity has a red, green, and blue component.
Similarly colorright is found by interpolating the colors at the top and bottom of the right edge. The
tiler then fills across the scanline, linearly interpolating between colorleft and colorright to obtain the
color at pixel x:

c( x ) = lerp(colorleft , colorright ,

x xleft
)
x right xleft

(8.14)

To increase efficiency this color is computed incrementally at each pixel. That is, there is a constant
difference between c(x+1) and c(x), so

c( x + 1) = c( x ) +

colorright colorleft
xright xleft

(8.15)

The increment is calculated only once outside of the innermost loop. In terms of code this looks like:
for (int y = ybott; y <= ytop; y++)
// for each scan-line
{
<.. find xleft and xright ..>
<.. find colorleft and colorright ..>
colorinc = (colorright colorleft)/(xright xleft);
for (int x = xleft, c = colorleft; x <= xright; x++, c+=colorinc)
<.. put c into the pixel at (x, y) ..>
}

We shall see later that, although colors are usually interpolated linearly as we do here, better results can be
obtained by using so-called hyperbolic interpolation. For Gouraud shading the distinction is minor; for texture
mapping it is crucial.
Chapter 8

November 30, 1999

page 20

Gouraud shading is modestly more expensive computationally than flat shading. Gouraud shading is
established in OpenGL using:
glShadeModel(GL_SMOOTH);
Figure 8.25 shows a buckyball and a sphere rendered using Gouraud shading. The buckyball looks
the same as when it was flat shaded in Figure 8.24, because the same color is associated with each
vertex of a face, so interpolation changes nothing. But the sphere looks much smoother. There are
no abrupt jumps in color between neighboring faces. The edges of the faces (and the Mach bands)
are gone, replaced by a smoothly varying color across the object. Along the silhouette, however,
you can still see the bounding edges of individual faces.

Figure 8.25. Two meshes rendered using smooth shading. (file: fig8.25.bmp)
Why do the edges disappear with this technique? Figure 8.26a shows two faces, F and F, that share
an edge. When rendering F the colors cL and cR are used, and when rendering F the colors cL and
cR are used. But since cR equals cL there is an abrupt change in color at the edge along the scanline.
a). two faces abuting
b). cross section: can see underlying surface
Figure 8.26. Continuity of color across a polygon edge.
Figure 8.26b suggests how this technique reveals the underlying surface approximated by the
mesh. The polygonal surface is shown in cross section, with vertices V1, V2, etc. marked. The
imaginary smooth surface that the mesh supposedly represents is suggested as well. Properly
computed vertex normals m1, m2, etc. point perpendicularly to this imaginary surface, so the normal
for correct shading is being used at each vertex, and the color thereby found is correct. The color
is then made to vary smoothly between vertices, not following any physical law but rather a simple
mathematical one.
Because colors are formed by interpolating rather than computing colors at every pixel, Gouraud
shading does not picture highlights well. Therefore, when Gouraud shading is used, one normally
suppresses the specular component of intensity in Equation 8.12. Highlights are better reproduced
using Phong shading, discussed next.
Phong Shading.
Greater realism can be achieved - particularly with regard to highlights on shiny objects - by a
better approximation of the normal vector to the face at each pixel. This type of shading is called
Phong shading, after its inventor Phong Bui-tuong [phong75].
When computing Phong shading we find the normal vector at each point on the face and we apply
the shading model there to find the color. We compute the normal vector at each pixel by
interpolating the normal vectors at the vertices of the polygon.
Figure 8.27 shows a projected face, with the normal vectors m1, m2, m3, and m4 indicated at the four
vertices. For the scan-line ys as shown the vectors mleft and mright are found by linear interpolation.
For instance, mleft is found as

Chapter 8

November 30, 1999

page 21

Figure 8.27. Interpolating normals.

m left = lerp(m 4 , m3 ,

ys y4
)
y3 y4

This interpolated vector must be normalized to unit length before its use in the shading formula.
Once mleft and mright are known, they are interpolated to form a normal vector at each x along the
scan-line. This vector, once normalized, is used in the shading calculation to form the color at that
pixel.
Figure 8.28 shows an object rendered using Gouraud shading and Phong shading. Because the
direction of the normal vector varies smoothly from point to point and more closely approximates
that of an underlying smooth surface, the production of specular highlights is much more faithful
than with Gouraud shading, and more realistic renderings are produced.
1st Ed. Figure 15.25
Figure 8.28. Comparison of Gouraud and Phong shading (Courtesy of Bishop and Weimar 1986).
The principal drawback of Phong shading is its speed: a great deal more computation is required per
pixel, so that Phong shading can take 6 to 8 times longer than Gouraud shading to perform. A
number of approaches have been taken to speed up the process [bishop86, claussen90].
OpenGL is not set up to do Phong shading, since it applies the shading model once per vertex right
after the modelview transformation, and normal vector information is not passed to the rendering
stage following the perspective transformation and perspective divide. We will see in Section 8.5,
however, that an approximation to Phong shading can be created by mapping a highlight texture
onto an object using the environment mapping technique.
Practice Exercises.
8.3.1. Filling your face. Fill in details of how the polygon fill algorithm operates for the polygon
with vertices (x, y) = (23, 137), (120, 204), (200, 100), (100, 25), for scan lines y = 136, y = 137,
and y = 138. Specifically write the values of xleft and xright in each case.
8.3.2. Clipped convex polygons are still convex. Develop a proof that if a convex polygon is
clipped against the cameras view volume, the clipped polygon is still convex.
8.3.3. Retaining edges with Gouraud Shading. In some cases we may want to show specific
creases and edges in the model. Discuss how this can be controlled by the choice of the vertex
normal vectors. For instance, to retain the edge between faces F and F in Figure 8.26, what should
the vertex normals be? Other tricks and issues can be found in the references [e.g. Rogers85].
8.3.4. Faster Phong shading with fence shading. To increase the speed of Phong shading Behrens
[behrens94] suggests interpolating normal vectors between vertices to get mL and mR in the usual
way at each scan line, but then computing colors only at these left and right pixels, interpolating
them along a scan line as in Gouraud shading. This so-called fence shading speeds up rendering
dramatically, but does less well in rendering highlights than true Phong shading. Describe general
directions for the vertex normals m1, m2, m3, and m4 in Figure 8.27 such that
a). Fence shading produces the same highlights as Phong shading;
b). Fence shading produces very different highlights than does Phong shading.
8.3.5. The Phong shading algorithm. Make the necessary changes to the tiling code to incorporate
Phong shading. Assume the vertex normal vectors are available for each face. Also discuss how
Phong shading can be approximated by OpenGLs smooth shading algorithm. Hint: increase the
number of faces in the model.

8.4. Adding Hidden Surface Removal

It is very simple to incorporate hidden surface removal in the rendering process above if enough
memory is available to have a depth buffer (also called a z-buffer). Because it fits so easily into
the rendering mechanisms we are discussing, we include it here. Other (more efficient and less
memory-hungry) hidden surface removal algorithms are described in Chapter 13.

8.4.1.The Depth Buffer Approach.

The depth buffer (or z-buffer) algorithm is one of the simplest and most easily implemented hidden
surface removal methods. Its principal limitations are that it requires a large amount of memory,

Chapter 8

November 30, 1999

page 22

and that it often renders an object that is later obscured by a nearer object (so time spent rendering
the first object is wasted).
Figure 8.29 shows a depth buffer associated with the frame buffer. For every pixel p[i][j] on the
display the depth buffer stores a b bit quantity d[i][j]. The value of b is usually in the range of 12 to
30 bits.

Figure 8.29. Conceptual view of the depth buffer.

During the rendering process the depth buffer value d[i][ j] contains the pseudodepth of the closest
object encountered (so far) at that pixel. As the tiler proceeds pixel by pixel across a scan-line
filling the current face, it tests whether the pseudodepth of the current face is less than the depth
d[i][j] stored in the depth buffer at that point. If so the color of the closer surface replaces the color
p[i][j] and this smaller pseudodepth replaces the old value in d[i][j]. Faces can be drawn in any
order. If a remote face is drawn first some of the pixels that show the face will later be replaced by
the colors of a nearer face. The time spent rendering the more remote face is therefore wasted. Note
that this algorithm works for objects of any shape including curved surfaces, because it finds the
closest surface based on a point-by-point test.
The array d[][] is initially loaded with value 1.0, the greatest pseudodepth value possible. The frame
buffer is initially loaded with the background color.
Finding the pseudodepth at each pixel.
We need a rapid way to compute the pseudodepth at each pixel. Recall that each vertex P = (Px, Py,
Pz) of a face is sent down the graphics pipeline, and passes through various transformations. The
information available for each vertex after the viewport transformation is the 3-tuple that is a scaled
and shifted version of (see Equation 7.2)

( x, y, z ) =

P , P , aP + b
P P P
y

The third component is pseudodepth. Constants a and b have been chosen so that the third component
equals 0 if P lies in the near plane, and 1 if P lies in the far plane. For highest efficiency we would like
to compute it at each pixel incrementally, which implies using linear interpolation as we did for color in
Equation 8.15.
Figure 8.30 shows a face being filled along scanline y. The pseudodepth values at various points are
marked. The pseudodepths d1, d2, d3, and d4 at the vertices are known. We want to calculate dleft at
scan-line ys as lerp(d1, d4, f) for fraction f = (ys y1)/(y4 y1) , and similarly dright as lerp(d2, d3, h) for
the appropriate h. And we want to find the pseudodepth d at each pixel (x, y) along the scan-line as
lerp(dleft, dright, k) for the appropriate k. (What are the values of h and k?) The question is whether
this calculation produces the true pseudodepth of the corresponding point on the 3D face.

Figure 8.30. Incremental computation of pseudodepth.

Figure 8.35. Drawing texture on several object shapes.

The rendering process actually goes the other way: for each pixel at (sx, sy) there is a sequence of
questions:
a). What is the closest surface seen at (sx, sy)? This determines which texture is relevant.
b). To what point (x, y, z) on this surface does (sx, sy) correspond?
c). To which texture coordinate pair (s, t) does this point (x, y, z) correspond?
So we need the inverse transformation, something like (s, t) = Ttw-1(Tws-1(sx, sy)), that reports (s, t)
coordinates given pixel coordinates. This inverse transformation can be hard to obtain or easy to
obtain, depending on the surface shapes.

8.5.1. Pasting the Texture onto a Flat Surface.

We first examine the most important case: mapping texture onto a flat surface. This is a
modeling task. In Section 8.5.2 we tackle the viewing task to see how the texture is actually
rendered. We then discuss mapping textures onto more complicated surface shapes.
Pasting Texture onto a Flat Face.
Since texture space itself is flat, it is simplest to paste texture onto a flat surface. Figure 8.36
shows a texture image mapped to a portion of a planar polygon F. We must specify how to
associate points on the texture with points on F. In OpenGL we associate a point in texture space
Pi = (si, ti) with each vertex Vi of the face using the function glTexCoord2f().The function
glTexCoord2f(s,t) sets the current texture coordinates to (s, t), and they are attached to
subsequently defined vertices. Normally each call to glVertex3f() is preceded by a call to
glTexCoord2f(), so each vertex gets a new pair of texture coordinates. For example, to
define a quadrilateral face and to position a texture on it, we send OpenGL four texture
coordinates and the four 3D points, as in:

Figure 8.36. Mapping texture onto a planar polygon.

glBegin(GL_QUADS); // define a quadrilateral face
Chapter 8

November 30, 1999

page 26

glTexCoord2f(0.0,
glTexCoord2f(0.0,
glTexCoord2f(0.8,
glTexCoord2f(0.8,
glEnd();

0.0);
0.6);
0.6);
0.0);

glVertex3f(1.0,
glVertex3f(1.0,
glVertex3f(2.0,
glVertex3f(2.0,

2.5,
3.7,
3.7,
2.5,

1.5);
1.5);
1.5);
1.5);

Attaching a Pi to each Vi is equivalent to prescribing a polygon P in texture space that has the
same number of vertices as F. Usually P has the same shape as F as well: then the portion of the
texture that lies inside P is pasted without distortion onto the whole of F. When P and F have
the same shape the mapping is clearly affine: it is a scaling, possibly accompanied by a rotation
and a translation.
Figure 8.37 shows the very common case where the four corners of the texture square are
associated with the four corners of a rectangle. (The texture coordinates (s, t) associated with
each corner are noted on the 3D face.) In this example the texture is a 640 by 480 pixel bitmap,
and it is pasted onto a rectangle with aspect ratio 640/480, so it appears without distortion. (Note
that the texture coordinates s and t still vary from 0 to 1.) Figure 8.38 shows the use of texture
coordinates that tile the texture, making it repeat. To do this some texture coordinates that lie
outside of the interval [0,1] are used. When the renderer encounters a value of s and t outside of
the unit square such as s = 2.67 it ignores the integral part and uses only the fractional part 0.67.
Thus the point on a face that requires (s, t) = (2.6, 3.77) is textured with texture(0.6, 0.77). By
default OpenGL tiles texture this way. It may be set to clamp texture values instead, if desired;
see the exercises.

Figure 8.37. Mapping a square to a rectangle.

Figure 8.38. Producing repeated textures.

Thus a coordinate pair (s, t) is sent down the pipeline along with each vertex of the face. As we
describe in the next section, the notion is that points inside F will be filled with texture values
lying inside P, by finding the internal coordinate values (s, t) using interpolation. This
interpolation process is described in the next section.
Adding texture coordinates to Mesh objects.
Recall from Figure 6.13 that a mesh object has three lists: the vertex, normal vector, and face
lists. We must add to this a texture coordinate list, that stores the coordinates (si, ti) to be
associated with various vertices. We can add an array of elements of the type:
class TxtrCoord{public: float s, t;};
to hold all of the coordinate pairs of interest for the mesh. There are several different ways to
treat texture for an object, and each has implications for how texture information is organized in
the model. The two most important are:
1.

The mesh object consists of a small number of flat faces, and a different texture is to be
applied to each. Here each face has only a single normal vector but its own list of texture
coordinates. So the data associated with each face would be:

the number of vertices in the face;

the index of the normal vector to the face;
a list of indices of the vertices;
a list of indices of the texture coordinates;

The mesh represents a smooth underlying object, and a single texture is to be wrapped
around it (or a portion of it). Here each vertex has associated with it a specific normal vector
and a particular texture coordinate pair. A single index into the vertex/normals/texture lists
is used for each vertex. The data associated with each face would then be:

Chapter 8

the number of vertices in the face;

November 30, 1999

page 27

a list of indices of the vertices;

The exercises take a further look at the required data structures for these types of meshes.

8.5.2. Rendering the Texture.

Rendering texture in a face F is similar to Gouraud shading: the renderer moves across the face
pixel by pixel. For each pixel it must determine the corresponding texture coordinates (s, t),
access the texture, and set the pixel to the proper texture color. We shall see that finding the
coordinates (s, t) must be done very carefully.
Figure 8.39 shows the camera taking a snapshot of face F with texture pasted onto it, and the
rendering in progress. Scanline y is being filled from xleft to xright. For each x along this scanline
we must compute the correct position (shown as P(x, y)) on the face, and from this obtain the
correct position (s*, t*) within the texture.

Figure 8.39. Rendering the face in a camera snapshot.

Having set up the texture to object mapping, we know the texture coordinates at each of the vertices
of F, as suggested in Figure 8.40. The natural thing is to compute (sleft, tleft) and (sright, tright) for each
scanline in a rapid incremental fashion and to interpolate between these values moving across the
scanline. But we must be careful: simple increments from sleft to sright as we march across scanline y
from xleft to xright wont work, since equal steps across a projected face do not correspond to equal
steps across the 3D face.

Figure 8.40. Incremental calculation of texture coordinates.

Figure 8.41 illustrates the problem. Part a shows face F viewed so that its left edge is closer to the
viewer than its right edge. Part b shows the projection F of this face on the screen. At scan-line y =
170 we mark points equally spaced across F, suggesting the positions of successive pixels on the
face. The corresponding positions of these marks on the actual face are shown in part a. They are
seen to be more closely spaced at the farther end of F. This is simply the effect of perspective
foreshortening.

Figure 8.41. Spacing of samples with linear interpolation.

If we use simple linear interpolation and take equally spaced steps in s and t to compute texture
coordinates, we sample into the texture at the wrong spots, and a distorted image results. Figure
8.42 shows what happens with a simple checkerboard texture mapped onto a rectangle. Linear
interpolation is used in part a, producing palpable distortion in the texture. This distortion is
particularly disturbing in an animation where the polygon is rotating, as the texture appears to warp
and stretch dynamically. Correct interpolation is used in part b, and the checkerboard looks as it
should. In an animation this texture would appear to be firmly attached to the moving or rotating
face.
a). linear
b). correct
Figure 8.42. Images formed using linear interpolation and correct interpolation.
Several approaches have appeared in the literature that develop the proper interpolation method.
Heckbert and Moreton [heckbert91] and Blinn [blinn92] describe an elegant development based on
the general nature of affine and projective mappings. Segal et al [segal92] arrive at the same result
using a more algebraic derivation based on the parametric representation for a line segment. We
follow the latter approach here.
Figure 8.43 shows the situation to be analyzed. We know that affine and projective
transformations preserve straightness, so line Le in eye space projects to line Ls in screen space,
and similarly the texels we wish to draw on line Ls lie along the line Lt in texture space that maps

Chapter 8

November 30, 1999

page 28

to Le. The key question is this: if we move in equal steps across Ls on the screen how should we
step across texels along Lt in texture space?

Figure 8.43. Lines in one space map to lines in another.

We develop a general result next that summarizes how interpolation works: it all has to do with
the effect of perspective division. Then we relate the general result to the transformations
performed in the graphics pipeline, and see precisely where extra steps must be taken to do proper
mapping of texture.
Figure 8.44 shows the line AB in 3D being transformed into the line ab in 3D by matrix M. (M
might represent an affine transformation, or a more general perpspective transformation.) A maps
to a, B maps to b. Consider the point R(g) that lies fraction g of the way between A and B. It
maps to some point r(f) that lies fraction f of the way from a to b. The fractions f and g are not the
same as we shall see. The question is, as f varies from 0 to 1 how exactly does g vary? That is,
how does motion along ab correspond to motion along AB?

Figure 8.44. How does motion along corresponding lines operate?

Deriving how g and f are related.
~ , and name its components
We denote the homogeneous coordinate version of a by a
~
a = ( a1 , a2 , a3 , a4 ) . (We use subscripts 1,2,3, and 4 instead of x, y, etc. to prevent ambiguity,
~ by perspective
since there are so many different x, y, z spaces.) So point a is found from a

a1 a2 a3
, , ) . Since M maps A= (A1, A2, A3) to a we know a~ = M ( A,1)T where (A,
a4 a4 a4
~
T
T
1) is the column vector with components A1, A2, A3, and 1. Similarly, b = M ( B,1) . (Check each
division: a = (

of these relations carefully.) Now using lerp() notation to keep things succinct, we have defined
T
~, b~, g)
R(g) = lerp(A, B, g), which maps to M (lerp( A, B, g),1) = lerp( a
= (lerp( a1 , b1 , g), lerp(a2 , b2 , g), lerp( a3 , b3 , g), lerp(a4 , b4 , g)) . (Check these, too.) This is
the homogeneous coordinate version ~
r ( f ) of the point r(f). We recover the actual components of
r(f) by perspective division. For simplicity write just the first component r1(f), which is:

r1 ( f ) =

lerp(a1 , b1 , g)
lerp(a4 , b4 , g)

(8.16)

But since by definition r(f) = lerp(a, b, f) we have another expression for the first component r1(f):

r1 ( f ) = lerp(

a1 b1
, , f)
a4 b4

(8.17)

Expressions (what are they?) for r2(f) and r3(f) follow similarly. Equate these two versions of r1(f)
and do a little algebra to obtain the desired relationship between f and g:

f
b4
lerp( ,1, f )
a4

(8.18)

Therefore the point R(g) maps to r(f), but g and f arent the same fraction. g matches at f = 0 and at f
= 1, but its growth with f is tempered by a denominator that depends on the ratio b4/a4. If a4 equals
b4 then g is identical to f (check this). Figure 8.45 shows how g varies with f , for different values of
b4/a4.
g vs f

Chapter 8

November 30, 1999

page 29

Figure 8.45. How g depends on f.

We can go the final step and show where the point R(g) is on the 3D face that maps into r(f).
Simply use Equation 8.17 in R(g) = A(1-g)+Bg and simplify algebraically (check this out) to obtain
for the first component:

A1 B1
, , f)
a4 b4
R1 =
1 1
lerp( , , f )
a4 b4
lerp(

(8.19)

with similar expressions resulting for the components R2 and R3 (which have the same denominator
as R1). This is a key result. It tells which 3D point (R1, R2, R3) corresponds (in eye coordinates) to a
given point that lies (fraction f of the way) between two given points a and b in screen coordinates.
So any quantity (such as texture) that is attached to vertices of the 3D face and varies linearly
between them will behave the same way.
The two cases of interest for the transformation with matrix M are:
The transformation is affine;
The transformation is the perspective transformation.
a).When the transformation is affine then a4 and b4 are both 1 (why?), so the formulas above
simplify immediately. The fractions f and g become identical, and R1 above becomes lerp(A1, B1, f).
We can summarize this as:
Fact: If M is affine, equal steps along the line ab do correspond to equal steps along the line AB.
b). When M represents the perspective transformation from eye coordinates to clip coordinates the
fourth components a4 and b4 are no longer 1.We developed the matrix M in Chapter 7. Its basic
form, given in Equation 7.10, is:

N
0
M=
0
0

0
N
0
0

0
0
c
1

d

0
0
0

where c and d are constants that make pseudodepth work properly. What is M(A,1)T for this matrix?
~ = ( NA , NA , cA + d, A ) , the crucial part being that a = -A . This is the position of the
Its a
4
3
1
2
3
3
point along the z-axis in camera coordinates, that is the depth of the point in front of the eye.
So the relative sizes of a4 and b4 lie at the heart of perspective foreshortening of a line segment: they
report the depths of A and B, respectively, along the cameras viewplane normal. If A and B have
the same depth (i.e. they lie in a plane parallel to the cameras viewplane), there is no perspective
distortion along the segment, so g and f are indeed the same. Figure 8.46 shows in cross section how
rays from the eye through evenly spaced spots (those with equal increments in f) on the viewplane
correspond to unevenly spaced spots on the original face in 3D. For the case shown A is closer than
B, causing a4 < b4, so the g-increments grow in size moving across the face from A to B.

Figure 8.46. The values of a4 and b4 are related to the depths of points.
Rendering incrementally.
We now put these ingredients together and find the proper texture coordinates (s, t) at each point on
the face being rendered. Figure 8.47 shows a face of the barn being rendered. The left edge of the
face has endpoints a and b. The face extends from xleft to xright across scan-line y. We need to find
appropriate texture coordinates (sleft, tleft) and (sright, tright) to attach to xleft and xright, respectively, which
we can then interpolate across the scan-line. Consider finding sleft(y), the value of sleft at scan-line y.
Chapter 8

November 30, 1999

page 30

We know that texture coordinate sA is attached to point a, and sB is attached to point b, since these
values have been passed down the pipeline along with the vertices A and B. If the scan-line at y is
fraction f of the way between ybott and ytop (so that f = (y ybott)/(ytop ybott)), then we know from
Equation 8.19 that the proper texture coordinate to use is:

Figure 8.47. Rendering the texture on a face.

sA sB
, , f)
a4 b4
sleft ( y ) =
1 1
lerp( , , f )
a4 b4
lerp(

(8.20)

and similarly for tleft. Notice that sleft and tleft have the same denominator: a linear interpolation
between values 1/a4 and 1/b4. The numerator terms are linear interpolations of texture coordinates
which have been divided by a4 and b4. This is sometimes called rational linear rendering
[heckbert91] or hyperbolic interpolation [blinn92]. To calculate (s, t) efficiently as f advances we
need to store values of sA/a4, sB/b4, tA/a4, tB/b4, 1/a4, and 1/b4, as these dont change from pixel to
pixel. Both the numerator and denominator terms can be found incrementally for each y, just as we
did for Gouraud shading (see Equation 8.15). But to find sleft and tleft we must still perform an explicit
division at each value of y.
The pair (sright, tright) is calculated in a similar fashion. They have denominators that are based on
values of a4 and b4 that arise from the projected points a and b.
Once (sleft, tleft) and (sright, tright) have been found the scan-line can be filled. For each x from xleft to xright
the values s and t are found, again by hyperbolic interpolation. (what is the expression for s at x?)
Implications for the graphics pipeline.
What are the implications of having to use hyperbolic interpolation to render texture properly? And
does the clipping step need any refinement? As we shall see, we must send certain additional
information down the pipeline, and calculate slightly different quantities than supposed so far.
Figure 8.48 shows a refinement of the pipeline. Various points are labeled with the information that
is available at that point. Each vertex V is associated with a texture pair (s, t) as well as a vertex
normal. The vertex is transformed by the modelview matrix (and the normal is multiplied by the
inverse transpose of this matrix), producing vertex A = (A1, A2, A3) and a normal n in eye
coordinates. Shading calculations are done using this normal, producing the color c = (cr, cg, cb). The
texture coordinates (sA, tA) (which are the same as (s, t)) are still attached to A. Vertex A then
~ = ( a , a , a , a ) . The texture coordinates
undergoes the perspective transformation, producing a
1
2
3
4
and color c are not altered.

Figure 8.48. Refinement of the graphics pipeline to include hyperbolic interpolation.

Now clipping against the view volume is done, as discussed in Chapter 7. As the figure suggests,
this can cause some vertices to disappear and others to be formed. When a vertex such as D is
created we must determine its position (d1, d2, d3, d4) and attach to it the appropriate color and
texture point. By the nature of the clipping algorithm the position components di are formed by
linear interpolation: di = lerp(ai, bi, t), for i = 1,.., 4, for some t. Notice that the fourth component d4
is also formed this way. It is natural to use linear interpolation here also to form both the color
components and the texture coordinates. (The rationale for this is discussed in the exercises.)
Therefore after clipping the face still consists of a number of vertices, and to each is attached a
color and a texture point. For point A the information is stored in the array (a1, a2, a3, a4, sA, tA, c, 1).
A final term of 1 has been appended: we will use it in the next step.
Now perspective division is done. Since for hyperbolic interpolation we need terms such as sA/a4
and 1/a4 (see Equation 8.20) we divide every item in the array that we wish to interpolate
hyperbolically by a4 to obtain the array (x, y, z, 1, sA/a4, tA/a4, c, 1/a4). (We could also divide the
color components in order to obtain slightly more realistic Gouraud shading. See the exercises.) The
Chapter 8

November 30, 1999

page 31

first three, (x, y, z) = (a1/a4, a2/a4, a3/a4) report the position of the point in normalized device
coordinates. The third component is pseudodepth. The first two components are scaled and shifted
by the viewport transformation. To simplify notation we shall continue to call the screen coordinate
point (x, y, z).
So finally the renderer receives the array (x, y, z, 1, sA/a4, tA/a4, c, 1/a4) for each vertex of the face to
be rendered. Now it is simple to render texture using hyperbolic interpolation as in Equation 8.20:
the required values sA/a4 and 1/a4 are available for each vertex.
Practice exercises.
8.5.1. Data structures for mesh models with textures. Discuss the specific data types needed to
represent mesh objects in the two cases:
a). a different texture is to be applied to each face;
b). a single texture is to be wrapped around the entire mesh.
Draw templates for the two data types required, and for each show example data in the various
arrays when the mesh holds a cube.
8.5.2. Pseudodepth calculations are correct. Show that it is correct, as claimed in Section 8.4, to
use linear (rather than hyperbolic) interpolation when finding pseudodepth. Assume point A projects
to a, and B projects to b. With linear interpolation we compute pseudodepth at the projected point
lerp(a, b, f) as the third component of this point. This is the correct thing to do only if the resulting
value equals the true pseudodepth of the point that lerp(A, B, g) (for the appropriate g) projects to.
Show that it is in fact correct. Hint: Apply Equations 8.16 and 8.17 to the third component of the
point being projected.
8.5.3. Wrapping and clamping textures in OpenGL. To make the pattern wrap or tile in
the s direction use: glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S,
GL_REPEAT). Similarly use GL_TEXTURE_WRAP_T for wrapping in the t-direction. This is
actually the default, so you neednt do this explicitly. To turn off tiling replace GL_REPEAT with
GL_CLAMP. Refer to the OpenGL documentation for more details, and experiment with different
OpenGL settings to see their effect.
8.5.4. Rationale for linear interpolation of texture during clipping. New vertices are often created
when a face is clipped against the view volume. We must assign texture coordinates to each vertex.
Suppose a new vertex V is formed that is fraction f of the way from vertex A to vertex B on a face.
Further suppose that A is assigned texture coordinates (sA, tA), and similarly for B. Argue why, if a
texture is considered as pasted onto a flat face, it makes sense to assign texture coordinates
(lerp(sA, sB, f), lerp(tA, tB, f)) to V.
8.5.5. Computational burden of hyperbolic interpolation. Compare the amount of computation
required to perform hyperbolic interpolation versus linear interpolation of texture coordinates.
Assume multiplication and division each require 10 times as much time as addition and subtraction.

8.5.3. What does the texture modulate?

How are the values in a texture map applied in the rendering calculation? We examine three
common ways to use such values in order to achieve different visual effects. We do it for the simple
case of the gray scale intensity calculation of Equation 8.5. For full color the same calculations are
applied individually for the red, green, and blue components.
1). Create a glowing object.
This is the simplest method computationally. The visible intensity I is set equal to the texture value
at each spot:

I = texture(s, t )
(or to some constant multiple of it). So the object appears to emit light or glow: lower texture values
emit less light and higher texture values emit more light. No additional lighting calculations need be
done.
(For colored light the red, green, and blue components are set separately: for instance, the red
component is Ir = texturer(s, t).)
To cause OpenGL to do this type of texturing, specify:

Chapter 8

November 30, 1999

page 32

glTexEnvf(GL_TEXUTRE_ENV,GL_TEXTURE_ENV_MODE, GL_REPLACE8);
2). Paint the texture by modulating the reflection coefficient.
We noted earlier that the color of an object is the color of its diffuse light component (when bathed
in white light). Therefore we can make the texture appear to be painted onto the surface by varying
the diffuse reflection coefficient, and perhaps the ambient reflection coefficient as well. We say that
the texture function modulates the value of the reflection coefficient from point to point. Thus we
replace Equation 8.5 with:

I = texture(s, t )[ Ia a + Id d lambert ] + Isp s phong f

for appropriate values of s and t. Since Phong specular reflections are the color of the source rather
than the object, highlights do not depend on the texture.
To cause OpenGL to do this type of texturing, specify:
glTexEnvf(GL_TEXTURE_ENV,GL_TEXTURE_ENV_MODE, GL_MODULATE);
3). Simulate roughness by Bump Mapping.
Bump mapping is a technique developed by Blinn [blinn78] to give a surface a wrinkled (like a
raisin) or dimpled (like an orange) appearance without struggling to model each dimple itself. Here
the texture function is used to perturb the surface normal vector, which causes perturbations in the
amount of diffuse and specular light. Figure 8.49 shows one example, and Plate ??? shows another.
One problem associated with bump mapping is that since the model itself does not contain the
dimples, the objects silhouette doesnt show dimples either, but is perfectly smooth along each
face. In addition, the corner between two adjacent faces is also visible in the silhouette. This can be
seen in the example.
screen shot bump mapping on a buckyball, showing some (smooth) edges
in silhouette
Figure 8.49. An apparently dimpled surface, achieved by bump mapping.
The goal is to make a scalar function texture(s, t) perturb the normal vector at each spot in a
controlled fashion. In addition, the perturbation should depend only on the surface shape and the
texture itself, and not on the orientation of the object or position of the eye. If it depended on
orientation, the dimples would change as the object moved in an animation, contrary to the desired
effect.
Figure 8.50 shows in cross section how bump mapping works. Suppose the surface is represented
parametrically by the function P(u, v), and has unit normal vector m(u, v). Suppose further that the
3D point at (u*, v*) corresponds to the texture at (u*, v*). Blinns method simulates perturbing the
position of the true surface in the direction of the normal vector by an amount proportional to
texture(u*, v*):
a).
b).
Figure 8.50. On the nature of bump mapping.
P(u*, v*) = P(u*, v*) + texture( u*, v*) m(u*, v*)

(8.21)

as shown in Figure 8.50a, which adds undulations and wrinkles in the surface. This perturbed
surface has a new normal vector m(u*, v*) at each point. The idea is to use this perturbed normal
as if it were attached to the original unperturbed surface at each point, as shown in Figure 8.50b.
Blinn shows that a good approximation to the m(u*, v*) (before normalization) is given by:
m(u*, v*) = m(u*, v*) + d(u*, v*)

(8.22)

where the perturbation vector d is given by

(u, v) = (m Pv )textureu (m Pu )texturev

Use either GL_REPLACE or GL_DECAL.

Chapter 8

November 30, 1999

page 33

where textureu and texturev are partial derivatives of the texture function with respect to u and v
respectively. Further, Pu and Pv are partial derivative of P(u, v) with respect to u and v,
respectively. All functions are evaluated at (u*, v*). Derivations of this result may also be
found in [watt2, miller98]. Note that the perturbation function depends only on the partial
derivatives of texture(), not on texture() itself.
If a mathematical expression is available for texture() you can form its partial derivatives
analytically. For example, texture() might undulate in two directions by combining sinewaves,
as in: texture(u, v) = sin(au)sin(bv) for some constants a and b. If the texture comes instead
from an image array, linear interpolation can be used to evaluate it at (u*, v*), and finite
differences can be used to approximate the partial derivatives.

8.5.4. A Texturing Example using OpenGL.

To illustrate how to invoke the texturing tools that OpenGL provides, we show an application that
displays a rotating cube having different images painted on its six sides. Figure 8.51 shows a
snapshot from the animation created by this program.

Figure 8.51. The textured cube generated by the example code.

The code for the application is shown in Figure 8.52. It uses a number of OpenGL functions to
establish the six textures and to attach them to the walls of the cube. There are many variations of
the parameters shown here that one could use to map textures. The version shown works well, but
careful adjustment of some parameters (using the OpenGL documentation as a guide) can improve
the images or increase performance. We discuss only the basics of the key routines.
One of the first tasks when adding texture to pictures is to create a pixmap of the texture in
memory. OpenGL uses textures that are stored in pixel maps, or pixmaps for short. These are
discussed in depth in Chapter 10, and the class RGBpixmap is developed that provides tools for
creating and manipulating pixmaps. Here we view a pixmap as a simple array of pixel values, each
pixel value being a triple of bytes to hold the red, green, and blue color values:
class RGB{ // holds a color triple each with 256 possible intensities
public: unsigned char r,g,b;
};
The RGBpixmap class stores the number of rows and columns in the pixmap, as well as the address of the
first pixel in memory:
class RGBpixmap{
public:
int nRows, nCols; // dimensions of the pixmap
RGB* pixel;
// array of pixels
int readBMPFile(char * fname); // read BMP file into this pixmap
void makeCheckerboard();
void setTexture(GLuint textureName);
};

Chapter 8

November 30, 1999

page 34

We show it here as having only three methods that we need for mapping textures. Other methods and details
are discussed in Chapter 10. The method readBMPFile() reads a BMP file9 and stores the pixel values in its
pixmap object; it is detailed in Appendix 3. The other two methods are discussed next.
Our example OpenGL application will use six textures. To create them we first make an RGBpixmap object
for each:
RGBpixmap pix[6];

// create six (empty) pixmaps

and then load the desired texture image into each one. Finally each one is passed to OpenGL to define a
texture.
1). Making a procedural texture.
We first create a checkerboard texture using the method makeCheckerboard(). The checkerboard
pattern is familiar and easy to create, and its geometric regularity makes it a good texture for testing
correctness. The application generates a checkerboard pixmap in pix[0] using:
pix[0].makeCheckerboard().
The method itself follows:
void RGBpixmap:: makeCheckerboard()
{ // make checkerboard patten
nRows = nCols = 64;
pixel = new RGB[3 * nRows * nCols];
if(!pixel){cout << out of memory!;return;}
long count = 0;
for(int i = 0; i < nRows; i++)
for(int j = 0; j < nCols; j++)
{
int c = (((i/8) + (j/8)) %2) * 255;
pixel[count].r = c;
// red
pixel[count].g = c;
// green
pixel[count++].b = 0;
// blue
}
}

It creates a 64 by 64 pixel array, where each pixel is an RGB triple. OpenGL requires that texture pixel maps
have a width and height that are both some power of two. The pixel map is laid out in memory as one long
array of bytes: row by row from bottom to top, left to right across a row. Here each pixel is loaded with the
value (c, c, 0), where c jumps back and forth between 0 and 255 every 8 pixels. (We used a similar jumping
method in Exercise 2.3.1.) The two colors of the checkerboard are black: (0,0,0), and yellow: (255,255,0). The
function returns the address of the first pixel of the pixmap, which is later passed to glTexImage2D() to
create the actual texture for OpenGL.
Once the pixel map has been formed, we must bind it to a unique integer name so that it can be referred to in
OpenGL without ambiguity. We arbitrarily assign the names 2001, 2002, , 2006 to our six textures in this
example11. The texture is created by making certain calls to OpenGL, which we encapsulate in the method:
void RGBpixmap :: setTexture(GLuint textureName)
{
glBindTexture(GL_TEXTURE_2D,textureName);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER,GL_NEAREST);
9

This is a standard device-independent image file format from Microsoft. Many images are available on the
internet in BMP format, and tools are readily available on the internet to convert other image formats to BMP
files.
10 A faster way that uses C++s bit manipulation operators is c = ((i&8)^(j&8))*255;
11 To avoid overlap in (integer) names in an application that uses many textures, it is better to let OpenGL
supply unique names for textures using glGenTextures(). If we need six unique names we can build an
array to hold them: GLuint name[6]; and then call glGenTextures(6,name). OpenGL places six
heretofore unused integers in name[0],,name[5], and we subsequently refer to the i-th texture using
name[i].
Chapter 8

November 30, 1999

page 35

glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB,nCols,nRows,0, GL_RGB,

GL_UNSIGNED_BYTE, pixel);
}
The call to glBindTexture() binds the given name to the texture being formed. When this call is made at a
later time, it will make this texture the active texture, as we shall see.
The calls to glTexParameteri()specify that a pixel should be filled with the texel whose coordinates are
nearest the center of the pixel, both when the texture needs to be magnified or reduced in size. This is fast but
can lead to aliasing effects. We discuss filtering of images and antialiasing further in Chapter 10. Finally, the
call to glTexImage2D() associates the pixmap with this current texture. This call describes the texture as
2D consisting of RGB byte-triples, gives its width, height, and the address in memory (pixel) of the first byte
of the bitmap.
2. Making a texture from a stored image.
OpenGL offers no support for reading an image file and creating the pixel map in memory. The
method readBMPFile() , given in Appendix 3, provides a simple way to read a BMP image into
a pixmap. For instance,
pix[1].readBMPFile("mandrill.bmp");
reads the file mandrill.bmp and creates the pixmap in pix[1].
Once the pixel map has been created, pix[1].setTexture() is used to pass the pixmap to
OpenGL to make a texture.
Texture mapping must also be enabled with glEnable(GL_TEXTURE_2D). In addition, the
routine glHint(GL_PERSPECTIVE_CORRECTION_HINT,GL_NICEST) is used to request
that OpenGL render the texture properly (using hyperbolic interpolation), so that it appears
correctly attached to faces even when a face rotates relative to the viewer in an animation.
// < the usual includes >
#include "RGBpixmap.h"
//######################## GLOBALS ########################
RGBpixmap pix[6];
// make six (empty) pixmaps
float xSpeed = 0, ySpeed = 0, xAngle = 0.0, yAngle = 0.0;
//<<<<<<<<<<<<<<<<<<<<<<<<<<< myinit >>>>>>>>>>>>>>>>>>>>>>>>>>>.
void myInit(void)
{
glClearColor(1.0f,1.0f,1.0f,1.0f); // background is white
glEnable(GL_DEPTH_TEST);
glEnable(GL_TEXTURE_2D);
pix[0].makeCheckerboard();
// make pixmap procedurally
pix[0].setTexture(2001);
// create texture
pix[1].readBMPFile("Mandrill.bmp"); // make pixmap from image
pix[1].setTexture(2002);
// create texture
//< similarly for other four textures >
glViewport(0, 0, 640, 480); // set up the viewing system
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluPerspective(60.0, 640.0/ 480, 1.0, 30.0); // set camera shape
glMatrixMode(GL_MODELVIEW);
glLoadIdentity();
glTranslated(0.0, 0.0, -4); // move camera back
}
//<<<<<<<<<<<<<<<<<<<<<<<<<<< display >>>>>>>>>>>>>>>>>>>>>>
void display(void)
{
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glTexEnvf(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_DECAL);
glPushMatrix();
glRotated(xAngle, 1.0,0.0,0.0); glRotated(yAngle, 0.0,1.0,0.0); //
rotate

Chapter 8

November 30, 1999

page 36

glBindTexture(GL_TEXTURE_2D,2001); // top face: 'fake' checkerboard

glBegin(GL_QUADS);
glTexCoord2f(-1.0, -1.0); glVertex3f(-1.0f, 1.0f, -1.0f);
glTexCoord2f(-1.0, 2.0); glVertex3f(-1.0f, 1.0f, 1.0f);
glTexCoord2f(2.0, 2.0); glVertex3f( 1.0f, 1.0f, 1.0f);
glTexCoord2f(2.0, -1.0); glVertex3f( 1.0f, 1.0f, -1.0f);
glEnd();
glBindTexture(GL_TEXTURE_2D,2002); // right face: mandrill
glBegin(GL_QUADS);
glTexCoord2f(0.0, 0.0); glVertex3f(1.0f, -1.0f, 1.0f);
glTexCoord2f(0.0, 2.0); glVertex3f(1.0f, -1.0f, -1.0f);
glTexCoord2f(2.0, 2.0); glVertex3f(1.0f, 1.0f, -1.0f);
glTexCoord2f(2.0, 0.0); glVertex3f(1.0f, 1.0f, 1.0f);
glEnd();
// < similarly for other four faces >
glFlush();
glPopMatrix();
glutSwapBuffers();
}
//<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<< spinner >>>>>>>>>>>>>>>>>>>>>
void spinner(void)
{ // alter angles by small amount
xAngle += xSpeed; yAngle += ySpeed;
display();
}
//<<<<<<<<<<<<<<<<<<<<<< main >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
void main(int argc, char **argv)
{
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_DOUBLE | GLUT_RGB | GLUT_DEPTH);
glutInitWindowSize(640,480);
glutInitWindowPosition(10, 10);
glutCreateWindow("rotating textured cube");
glutDisplayFunc(display);
myInit();
glutIdleFunc(spinner);
glutMainLoop();
}
Figure 8.52. An application of a rotating textured cube.
The texture creation, enabling, and hinting needs to be done only once, in an initialization routine. Then
each time through the display routine the texture is actually applied. In display() the cube is rotated
through angles xAngle, and yAngle, and the six faces are drawn. This requires simply that the
appropriate texture be bound to the face and that within a glBegin()/glEnd() pair the texture coordinates
and 3D positions of the faces vertices be specified, as shown in the code.
Once the rendering (off screen) of the cube is complete glutSwapBuffers() is called to make the new
frame visible. The animation is controlled by using the callback function spinner() as the idle function.
Whenever the system is idle not responding to user input spinner is called automatically. It alters the
rotation angles of the cube slightly, and calls display() once again. The effect is an ongoing animation
showing the cube rotating, so that its various faces come into view and rotate out of view again and again.

8.5.5. Wrapping texture on curved surfaces.

We have seen how to paste a texture onto a flat surface. Now we examine how to wrap texture
onto a curved surface, such as a beer can or a chess piece. We assume as before that the object is
modeled by a mesh, so it consists of a large number of small flat faces. As discussed at the end
of Section 8.5.1 each vertex of the mesh has an associated texture coordinate pair (si, ti). The
main question is finding the proper texture coordinate (s, t) for each vertex of the mesh.
We present examples of mapping textures onto cylinder-like objects and sphere-like objects,
and see how a modeler might deal with each one.
Example 8.5.1. Wrapping a label around a can.
Chapter 8

November 30, 1999

page 37

Suppose that we want to wrap a label about a circular cylinder, as suggested in Figure 8.53a. Its
natural to think in terms of cylindrical coordinates. The label is to extend from a to b in
azimuth and from za to zb along the z-axis. The cylinder is modeled as a polygonal mesh, so its
walls are rectangular strips as shown in part b. For vertex Vi of each face we must find suitable
texture coordinates (si, ti), so that the correct slice of the texture is mapped onto the face.
a).
b).
Figure 8.53. Wrapping a label around a cylinder.
The geometry is simple enough here that a solution is straightforward. There is a direct linear
relationship between (s, t) and the azimuth and height (, z) of a point on the cylinders surface:

a
z za
, t=
b a
zb za

(8.23)

So if there are N faces around the cylinder, the i-th face has left edge at azimuth i = 2i/N, and
its upper left vertex has texture coordinates (si, ti) = ((2i/N - a)/(b-a), 1). Texture coordinates
for the other three vertices follow in a similar fashion. This association between (s, t) and the
vertices of each face is easily put in a loop in the modeling routine (see the exercises).
Things get more complicated when the object isnt a simple cylinder. We see next how to map
texture onto a more general surface of revolution.
Example 8.5.2. Shrink wrapping a label onto a Surface of Revolution.
Recall from Chapter 6 that a surface of revolution is defined by a profile curve (x(v), z(v))12 as
shown in Figure 8.54a, and the resulting surface - here a vase - is given parametrically by P(u, v)
= (x(v) cos u, x(v) sin u, z(v)). The shape is modeled as a collection of faces with sides along
contours of constant u and v (see Figure 8.54b). So a given face Fi has four vertices P(ui, vi),
P(ui+1, vi), P(ui, vi+1), and P(ui+1, vi+1). We need to find the appropriate (s, t) coordinates for each of
these vertices.
a). a vase profile
b). a face on the vase - four corners
Figure 8.54.Wrapping a label around a vase.
One natural approach is to proceed as above and to make s and t vary linearly with u and v in the
manner of Equation 8.23. This is equivalent to wrapping the texture about an imaginary rubber
cylinder that encloses the vase (see Figure 8.55a), and then letting the cylinder collapse, so that
each texture point slides radially (and horizontally) until it hits the surface of the vase. This
method is called shrink wrapping by Bier and Sloane [bier86], who discuss several possible
ways to map texture onto different classes of shapes. They view shrink wrapping in terms of the
imaginary cylinders normal vector (see Figure 8.55b): texture point Pi is associated with the
object point Vi that lies along the normal from Pi.

Figure 8.55. Shrink wrapping texture onto the vase.

Shrink wrapping works well for cylinder-like objects, although the texture pattern will be
distorted if the profile curve has a complicated shape.
Bier and Sloane suggest some alternate ways to associate texture points on the imaginary
cylinder with vertices on the object. Figure 8.56 shows two other possibilities.
a). centroid
b). object normal
Figure 8.56. Alternative mappings from the imaginary cylinder to the object.

We revert to calling the parameters u and v in the parametric representation of the shape, since we are using
s and t for the texture coordinates.

Chapter 8

November 30, 1999

page 38

In part a) a line is drawn from the objects centroid C, through the vertex Vi, to its intersection
with the cylinder Pi. And in part b) the normal vector to the objects surface at Vi is used: Pi is at
the intersection of this normal from Vi with the cylinder. Notice that these three ways to
associate texture points with object points can lead to very different results depending on the
shape of the object (see the exercises). The designer must choose the most suitable method based
on the objects shape and the nature of the texture image being mapped. (What would be
appropriate for a chess pawn?)
Example 8.5.3. Mapping texture onto a sphere.
It was easy to wrap a texture rectangle around a cylinder: topologically a cylinder can be sliced
open and laid flat without distortion. A sphere is a different matter. As all map makers know,
there is no way to show accurate details of the entire globe on a flat piece of paper: if you slice
open a sphere and lay it flat some parts always suffer serious stretching. (Try to imagine a
checkerboard mapped over an entire sphere!)
Its not hard to paste a rectangular texture image onto a portion of a sphere, however. To map
the texture square to the portion lying between azimuth a to b and latitude a to b just map
linearly as in Equation 8.23: if vertex Vi lies a (i, i) associate it with texture coordinates (si, ti) =
((i - a)/((b - a), (i - a)/(b - a)). Figure 8.57 shows an image pasted onto a band around a
sphere. Only a small amount of distortion is seen.
a). texture on portion of sphere
b). 8 maps onto 8 octants
Figure 8.57. Mapping texture onto a sphere.
Figure 8.57b shows how one might cover an entire sphere with texture: map eight triangular
texture maps onto the eight octants of the sphere.
Example 8.5.4. Mapping texture to sphere-like objects.
We discussed adding texture to cylinder-like objects above. But some objects are more spherelike than cylinder-like. Figure 8.58a shows the buckyball, whose faces are pentagons and
hexagons. One could devise a number of pentagonal and hexagonal textures and manually paste
one of each face, but for some scenes it may be desirable to wrap the whole buckyball in a single
texture.
a). buckyball
b). three mapping methods
Figure 8.58. Sphere-like objects.
It is natural to surround a sphere-like object with an imaginary sphere (rather than a cylinder)
that has texture pasted to it, and use one of the association methods discussed above. Figure
8.58b shows the buckyball surrounded by such a sphere in cross section. The three ways of
associating texture points Pi with object vertices Vi are sketched:
object-centroid: Pi is on a line from the centroid C through vertex Vi;
object-normal: Pi is the intersection of a ray from Vi in the direction of the face normal;
sphere-normal: Vi is the intersection of a ray from Pi in the direction of the normal to the sphere
at Pi.
(Question: Are the object-centroid and sphere-normal methods the same if the centroid of the
object coincides with the center of the sphere?) The object centroid method is most likely the
best, and it is easy to implement. As Bier and Sloane argue, the other two methods usually
produce unacceptable final renderings.
Bier and Sloane also discuss using an imaginary box rather than a sphere to surround the object
in question. Figure 8.59a shows the six faces of a cube spread out over a texture image, and part
b) shows the texture wrapped about the cube, which in turn encloses an object. Vertices on the
object can be associated with texture points in the three ways discussed above: the objectcentroid and cube-normal are probably the best choices.
a). texture on 6 faces of box
b). wrapping texture onto
Figure 8.59. Using an enclosing box.
Chapter 8

November 30, 1999

page 39

Practice exercises.
8.5.7. How to associate Pi and Vi. Surface of revolution S shown in Figure 8.60 consists of a
sphere resting on a cylinder. The object is surrounded by an imaginary cylinder having a
checkerboard texture pasted on it. Sketch how the texture will look for each of the following
methods of associating texture points to vertices:
a). shrink wrapping;
b). object centroid;
c). object normal;

Figure 8.60. A surface of revolution surrounded by an imaginary cylinder.

8.5.8. Wrap a texture onto a torus. A torus can be viewed as a cylinder that bends around
and closes on itself. The torus shown in Figure 8.61 has the parametric representation given by
P(u, v) = ((D + A cos(v)) cos(u), (D + A cos(v)) sin(u), A sin(v)). Suppose you decide to
polygonalize the torus by taking vertices based on the samples ui = 2i/N and vj = 2j/M, and you
wish to wrap some texture from the unit texture space around this torus. Write code that
generates, for each of the faces, each vertex and its associated texture coordinates (s, t).

Figure 8.61. Wrapping texture about a torus.

8.5.6. Reflection mapping.

The class of techniques known as reflection mapping can significantly improve the realism of
pictures, particularly in animations. The basic idea is to see reflections in an object that suggest the
world surrounding that object.
The two main types of reflection mapping are called chrome mapping and environment
mapping. In the case of chrome mapping a rough and usually blurry image that suggests the
surrounding environment is reflected in the object, as you would see in a surface coated with
chrome. Television commercials abound with animations of shiny letters and logos flying around in
space, where the chrome map includes occasional spotlights for dramatic effect. Figure 8.62 offers
an example. Part a) shows the chrome texture, and part b) shows it reflecting in the shiny object.
The reflection provides a rough suggestion of the world surrounding the object.
a). chrome map
b). scene with chrome mapping
(screen shots)
Figure 8.62. Example of chrome mapping.
In the case of environment mapping (first introduced by Blinn and Newell [blinn 76]) a
recognizable image of the surrounding environment is seen reflected in the object. We get valuable
visual cues from such reflections, particularly when the object moves about. Everyone has seen the
classic photographs of an astronaut walking on the moon with the moonscape reflected in his face
mask. And in the movies you sometimes see close-ups of a character's reflective dark glasses, in
which the world about her is reflected. Figure 8.63 shows two examples where a cafeteria is
reflected in a sphere and a torus. The cafeteria texture is wrapped about a large sphere that
surrounds the object, so that the texture coordinates (s, t) correspond to azimuth and latitude about
the enclosing sphere.

Figure 8.63. Example of environment mapping (courtesy of Haeberli and Segal).

Chapter 8

November 30, 1999

page 40

Figure 8.64 shows the use of a surrounding cube rather than a sphere. Part a) shows the map,
consisting of six images of various views of the interior walls, floor, and ceiling of a room. Part b)
shows a shiny object reflecting different parts of the room. The use of an enclosing cube was
introduced by Greene [greene 86], and generally produces less distorted reflections than are seen
with an enclosing sphere. The six maps can be generated by rendering six separate images from the
point of view of the object (with the object itself removed, of course). For each image a synthetic
camera is set up and the appropriate window is set. Alternatively, the textures can be digitized from
photos taken by a real camera that looks in the six principal directions inside an actual room or
scene.
a). six images make the map
b). environment mapping
(screen shots)
Figure 8.64. Environment mapping based on a surrounding cube.
Chrome and environment mapping differ most dramatically from normal texture mapping in an
animation when the shiny object is moving. The reflected image will flow over the moving
object, whereas a normal texture map will be attached to the object and move with it. And if a shiny
sphere rotates about a fixed spot a normal texture map spins with the sphere, but a reflection map
stays fixed.
How is environment mapping done? What you see at point P on the shiny object is what has arrived
at P from the environment in just the right direction to reflect into your eye. To find that direction
trace a ray from the eye to P, and determine the direction of the reflected ray. Trace this ray to find
where it hits the texture (on the enclosing cube or sphere). Figure 8.65 shows a ray emanating from
the eye to point P. If the direction of this ray is u and the unit normal at P is m, we know from
Equation 8.2 that the reflected ray has direction r = u 2(u m)m. The reflected ray moves in
direction r until it hits the hypothetical surface with its attached texture. It is easiest
computationally to suppose that the shiny object is centered in, and much smaller than, the
enclosing cube or sphere. Then the reflected ray emanates approximately from the objects center,
and its direction r can be used directly to index into the texture.

Figure 8.65. Finding the direction of the reflected ray.

OpenGL provides a tool to perform approximate environment mapping for the case where the
texture is wrapped about a large enclosing sphere. It is invoked by setting a mapping mode for both
s and t using:
glTexGenf(GL_S,GL_TEXTURE_GEN_MODE, GL_SPHERE_MAP);
glTexGenf(GL_T,GL_TEXTURE_GEN_MODE, GL_SPHERE_MAP);
glEnable(GL_TEXTURE_GEN_S);
glEnable(GL_TEXTURE_GEN_T);
Now when a vertex P with its unit normal m is sent down the pipeline, OpenGL calculates a texture
coordinate pair (s, t) suitable for indexing into the texture attached to the surrounding sphere. This
is done for each vertex of the face on the object, and the face is drawn as always using interpolated
texture coordinates (s, t) for points in between the vertices.
How does OpenGL rapidly compute a suitable coordinate pair (s, t)? As shown in Figure 8.66a it
first finds (in eye coordinates) the reflected direction r (using the formula above), where u is the
unit vector (in eye coordinates) from the eye to the vertex V on the object, and m is the normal at V.
a).
b).
Figure 8.66. OpenGLs computation of the texture coordinates.
It then simply uses the expression:

(s, t ) =

Chapter 8

1
2

ry
rx
+ 1), 12 ( + 1)
p
p

(8.24)

November 30, 1999

page 41

where p is a mysterious scaling factor p =

rx 2 + ry 2 + (rz + 1)2 . The derivation of this term is

developed in the exercises. We must precompute a texture that shows what you would see of the
environment in a perfectly reflecting sphere, from an eye position far removed from the sphere
[haeberli93]. This maps the part of the environment that lies in the hemisphere behind the eye into a
circle in the middle of the texture, and the part of the environment in the hemisphere in front of the
eye into an annulus around this circle (visualize this). This texture must be recomputed if the eye
changes position. The pictures in Figure 8.63 were made using this method.
Simulating Highlights using Environment mapping.
Reflection mapping can be used in OpenGL to produce specular highlights on a surface. A texture
map is created that has an intense concentrated bright spot. Reflection mapping paints this
highlight onto the surface, making it appear to be an actual light source situated in the environment.
The highlight created can be more concentrated and detailed than those created using the Phong
specular term with Gouraud shading. Recall that the Phong term is computed only at the vertices of
a face, and it is easy to miss a specular highlight that falls between two vertices. With reflection
mapping the coordinates (s, t) into the texture are formed at each vertex, and then interpolated in
between. So if the coordinates indexed by the vertices happen to surround the bright spot, the spot
will be properly rendered inside the face.
Practice Exercise 8.5.9. OpenGLs computation of texture coordinates for environment
mapping. Derive the result in Equation 8.24. Figure 8.66b shows in cross-sectional view the vectors
involved (in eye coordinates). The eye is looking from a remote location in the direction (0,0,1). A
sphere of radius 1 is positioned on the negative z-axis. Suppose light comes in from direction r,
hitting the sphere at the point (x, y, z). The normal to the sphere at this point is (x, y, z), which also
must be just right so that light coming along r is reflected into the direction (0, 0, 1). This means
the normal must be half-way between r and (0, 0, 1), or must be proportional to their sum, so (x, y,
z) = K(rx, ry, rz+1) for some K.
a). Show that the normal vector has unit length if K is 1/p, where p is given as in Equation 8.24.
b). Show that therefore (x, y) = (rx/p, ry/p).
c). Suppose for the moment that the texture image extends from 1 to 1 in x and from 1 to 1 in y.
Argue why what we want to see reflected at the point (x, y, z) is the value of the texture image at (x,
y). d). Show that if instead the texture uses coordinates from 0 to 1 as is true with OpenGL that
we want to see at (x, y) the value of the texture image at (s, t) given by Equation 8.24.

8.6. Adding Shadows of Objects.

Shadows make an image much more realistic. From everyday experience the way one object casts a
shadow on another object gives important visual cues as to how they are positioned. Figure 8.67
shows two images involving a cube and a sphere suspended above a plane. Shadows are absent in
part a, and it is impossible to see how far above the plane the cube and sphere are floating. By
contrast, the shadows seen in part b give useful hints as to the positions of the objects. A shadow
conveys a lot of information; its as if you are getting a second look at the object (from the
viewpoint of the light source).
a). with no shadows
b). with shadows
(screen shots)
Figure 8.67. The effect on shadows.
In this section we examine two methods for computing shadows: one is based on painting
shadows as if they were texture, and the other is an adaptation of the depth buffer approach for
hidden surface removal. In Chapter 14 we see that a third method arises naturally when raytracing.
There are many other techniques, well surveyed in [watt92, crow77, woo90, bergeron86].

8.6.1. Shadows as Texture.

This technique displays shadows that are cast onto a flat surface by a point light source. The
problem is to compute the shape of the shadow that is cast. Figure 8.68a shows a box casting a
shadow onto the floor. The shape of the shadow is determined by the projections of each of the
faces of the box onto the plane of the floor, using the source as the center of projection. In fact the
shadow is the union13 of the projections of the six faces. Figure 8.68b shows the superposed

the set theoretic union: A point is in the shadow if it is in one or more of the projections.

Chapter 8

November 30, 1999

page 42

projections of two of the faces: the top face projects to top and the front face to front. (Sketch the
projections of the other four faces, and see that their union is the required shadow14.)
a).
b).
Figure 8.68. Computing the shape of a shadow.
This is the key to drawing the shadow. After drawing the plane using ambient, diffuse, and specular
light contributions, draw the six projections of the boxs faces on the plane using only ambient light.
This will draw the shadow in the right shape and color. Finally draw the box. (If the box is near the
plane parts of it might obscure portions of the shadow.)
Building the projected face:
To make the new face F produced by F, project each of its vertices onto the plane in question. We
need a way to calculate these vertex positions on the plane. Suppose, as in Figure 8.68a, that the
plane passes through point A and has normal vector n. Consider projecting vertex V, producing
point V. The mathematics here are familiar: Point V is the point where the ray from the source at S
through V hits the plane. As developed in the exercises, this point is:

V ' = S + (V - S )

n ( A - S)
n (V - S )

(8.25)

The exercises show how this can be written in homogeneous coordinates as V times a matrix, which
is handy for rendering engines, like OpenGL, that support convenient matrix multiplication.
Practice Exercises.
8.6.1. Shadow shapes. Suppose a cube is floating above a plane. What is the shape of the cubes
shadow if the point source lies a). directly above the top face? b). along a main diagonal of the cube
(as in an isometric view)? Sketch shadows for a sphere and for a cylinder floating above a plane for
various source positions.
8.6.2. Making the shadow face. a). Show that the ray from the source point S through vertex V
hits the plane n ( P A) = 0 at t* = n ( A - S ) / n (V - S ) ; b). Show that this defines the hit
point V as given in Equation 8.25.
8.6.3. Its equivalent to a matrix multiplication. a). Show that the expression for V in Equation
T
8.25 can be written as a matrix multiplication: V ' = M (Vx , Vy , Vz ,1) , where M is a 4 by 4 matrix
b). Express the terms of M in terms of A, S, and n.

8.6.2. Shadows using a shadow buffer.

A rather different method for drawing shadows uses a variant of the depth buffer that performs
hidden surface removal. It uses an auxiliary second depth buffer, called a shadow buffer, for each
light source. This requires a lot of memory, but this approach is not restricted to casting shadows
onto planar surfaces.
The method is based on the principle that any points in the scene that are hidden from the light
source must be in shadow. On the other hand, if no object lies between a point and the light source
the point is not in shadow. The shadow buffer contains a depth picture of the scene from the point
of view of the light source: each of its elements records the distance from the source to the closest
object in the associated direction.
Rendering is done in two stages:
1). Shadow buffer loading. The shadow buffer is first initialized with 1.0 in each element, the
largest pseudodepth possible. Then, using a camera positioned at the light source, each of the faces
in the scene is scan converted, but only the pseudodepth of the point on the face is tested. Each
element of the shadow buffer keeps track of the smallest pseudodepth seen so far.
To be more specific, Figure 8.69 shows a scene being viewed by the usual eye camera as well as
a source camera located at the light source. Suppose point P is on the ray from the source through
14

You need to form the union of the projections of only the three front faces: those facing toward the light
source. (Why?)

Chapter 8

November 30, 1999

page 43

shadow buffer pixel d[i][j], and that point B on the pyramid is also on this ray. If the pyramid is
present d[i][j] contains the pseudodepth to B; if it happens to be absent d[i][j] contains the
pseudodepth to P.

Figure 8.69. Using the shadow buffer.

Note that the shadow buffer calculation is independent of the eye position, so in an animation where
only the eye moves the shadow buffer is loaded only once. The shadow buffer must be recalculated,
however, whenever the objects move relative to the light source.
2). Render the scene. Each face in the scene is rendered using the eye camera as usual. Suppose the
eye camera sees point P through pixel p[c][r]. When rendering p[c][r] we must find15:

the pseudodepth D from the source to P;

the index location [i][j] in the shadow buffer that is to be tested;
the value d[i][j] stored in the shadow buffer.

If d[i][j] is less than D the point P is in shadow, and p[c][r] is set using only ambient light.
Otherwise P is not in shadow and p[c][r] is set using ambient, diffuse, and specular light.
How are these steps done? As described in the exercises, to each point on the eye camera viewplane
there corresponds a point on the source camera viewplane16. For each screen pixel this
correspondence is invoked to find the pseudodepth from the source to P as well as the index [i][j]
that yields the minimum pseudodepth stored in the shadow buffer.
Practice Exercises.
8.6.4. Finding pseudodepth from the source. Suppose the matrices Mc and Ms map the point P in
the scene to the appropriate (3D) spots on the eye cameras viewplane and the source cameras
viewplane, respectively. a). Describe how to establish a source camera and how to find the
resulting matrix Ms. b). Find the transformation that, given position (x, y) on the eye cameras
viewplane produces the position (i, j) and pseudodepth on the source cameras viewplane.
c). Once (i, j) are known, how is the index [i][j] and the pseudodepth of P on the source camera
determined?
8.6.5. Extended Light sources. We have considered only point light sources in this chapter.
Greater realism is provided by modeling extended light sources. As suggested in Figure 8.70a such
sources cast more complicated shadows, having an umbra within which no light from the source is
seen, and a lighter penumbra within which a part of the source is visible. In part b) a glowing
sphere of radius 2 shines light on a unit cube, thereby casting a shadow on the wall W. Make an
accurate sketch of the umbra and penumbra that is observed on the wall. As you might expect,
algorithms for rendering shadows due to extended light sources are complex. See [watt92] for a
thorough treatment.
a). umbra and penumbra
b). example to sketch
Figure 8.70. Umbra and penumbra for extended light sources.

8.7. Summary
Since the beginning of computer graphics there has been a relentless quest for greater realism
when rendering 3D scenes. Wireframe views of objects can be drawn very rapidly but are
difficult to interpret, particularly if several objects in a scene overlap. Realism is greatly
enhanced when the faces are filled with some color and surfaces that should be hidden are
removed, but pictures rendered this way still do not give the impression of objects residing in a
scene, illuminated by light sources.

Of course, this test is made only if P is closer to the eye than the value stored in the normal depth
buffer of the eye camera.

Keep in mind these are 3D points: 2 position coordinates on the viewplane, and pseudodepth.

Chapter 8

November 30, 1999

page 44

What is needed is a shading model, that describes how light reflects off a surface depending on
the nature of the surface and its orientation to both light sources and the cameras eye. The
physics of light reflection is very complex, so programmers have developed a number of
approximations and tricks that do an acceptable job most of the time, and are reasonably
efficient computationally. The model for the diffuse component is the one most closely based
on reality, and becomes extremely complex as more and more ingredients are considered.
Specular reflections are not modeled on physical principles at all, but can do an adequate job of
recreating highlights on shiny objects. And ambient light is purely an abstraction, a shortcut
that avoids dealing with multiple reflections from object to object, and prevents shadows from
being too deep.
Even simple shading models involve several parameters such as reflection coefficients,
descriptions of a surfaces roughness, and the color of light sources. OpenGL provides ways to
set many of these parameters. There is little guidance for the designer in choosing the values of
these parameters; they are often determined by trial and error until the final rendered picture
looks right.
In this chapter we focused on rendering of polygonal mesh models, so the basic task was to
render a polygon. Polygonal faces are particularly simple and are described by a modest
amount of data, such as vertex positions, vertex normals, surface colors and material. In
addition there are highly efficient algorithms for filling a polygonal face with calculated colors,
especially if it is known to be convex. And algorithms can capitalize on the flatness of a
polygon to interpolate depth in an incremental fashion, making the depth buffer hidden surface
removal algorithm simple and efficient.
When a mesh model is supposed to approximate an underlying smooth surface the appearance
of a faces edges can be objectionable. Gouraud and Phong shading provide ways to draw a
smoothed version of the surface (except along silhouettes). Gouraud shading is very fast but
does not reproduce highlights very faithfully; Phong shading produces more realistic renderings
but is computationally quite expensive.
The realism of a rendered scene is greatly enhanced by the appearance of texturing on object
surfaces. Texturing can make an object appear to be made of some material such as brick or
wood, and labels or other figures can be pasted onto surfaces. Texture maps can be used to
modulate the amount of light that reflects from an object, or as bump maps that give a
surface a bumpy appearance. Environment mapping shows the viewer an impression of the
environment that surrounds a shiny object, and this can make scenes more realistic, particularly
in animations. Texture mapping must be done with care, however, using proper interpolation
and antialiasing (as we discuss in Chapter 10).
The chapter closed with a description of some simple methods for producing shadows of
objects. This is a complex subject, and many techniques have been developed. The two
algorithms described provide simple but partial solutions to the problem.
Greater realism can be attained with more elaborate techniques such as ray tracing and
radiosity. Chapter 14 develops the key ideas of these techniques.

8.8. Case Studies.

8.8.1. Case Study 8.1. Creating shaded objects using OpenGL
(Level of Effort: II beyond that of Case Study 7.1). Extend Case Study 7.1 that flies a camera
through space looking at various polygonal mesh objects. Extend it by establishing a point light
source in the scene, and assigning various material properties to the meshes. Include ambient,
diffuse, and specular light components. Provide a keystroke that switches between flat and
smooth shading

8.8.2. Case Study 8.2. The Do-it-yourself graphics pipeline.

(Level of Effort: III) Write an application that reads a polygonal mesh model from a file as
described in Chapter 6, defines a camera and a point light source, and renders the mesh object
using flat shading with ambient and diffuse light contributions. Only gray scale intensities need
be computed. For this project do not use OpenGLs pipeline; instead create your own. Define
modelview, perspective, and viewport matrices. Arrange that vertices can be passed through the
Chapter 8

November 30, 1999

page 45

first two matrices, have the shading model applied, followed by perspective division (no
clipping need be done) and by the viewport transformation. Each vertex emerges as the array
{x, y, z, b} where x and y are screen coordinates, z is pseudodepth, and b is the grayscale
brightness of the vertex. Use a tool that draws filled polygons to do the actual rendering: if you
use OpenGL, use only its 2D drawing (and depth buffer) components. Experiment with
different mesh models, camera positions, and light sources to insure that lighting is done
properly.

8.8.3. Case Study 8.3. Add Polygon Fill and Depth Buffer HSR.
(Level of Effort: III beyond that needed for Case Study 8.2.) Implement your own depth buffer,
and use it in the application of Case Study 8.2. This requires the development of a polygon fill
routine as well - see Chapter 10.

8.8.4. Case Study 8.4. Texture Rendering.

(Level of Effort: II beyond that of Case Study 8.1). Enhance the program of Case Study 8.1 so
that textures can be painted on the faces of the mesh objects. Assemble a routine that can read a
BMP image file and attach it to an OpenGL texture object. Experiment by putting five different
image textures and one procedural texture on the sides of a cube, and arranging to have the
cube rotate in an animation. Provide a keystroke that lets the user switch between linear
interpolation and correct interpolation for rendering textures.

8.8.5. Case Study 8.5. Applying Procedural 3D textures.

(Level of Effort: III) An interesting effect is achieved by making an object appear to be carved
out of some solid material, such as wood or marble. Plate ??? shows a (raytraced) vase carved
out of marble, and Plate ??? shows a box apparently made of wood. 3D textures are discussed
in detail in Chapter 14 in connection with ray tracing, but it is also possible to map slices of a
3D texture onto the surfaces of an object, to achieve a convincing effect.
Suppose you have a texture function B(x, y, z) which attaches different intensities or colors to
different points in 3D space. For instance B(x, y, z) might represent how inky the sea is at
position (x, y, z). As you swim around you encounter a varying inkiness right before your eyes. If
you freeze a block of this water and carve some shape out of the block, the surface of the shape
will exhibit a varying inkiness. B() can be vector-valued as well: providing three values at each
(x, y, z), which might represent the diffuse reflection coefficients for red, green, and blue light of
the material at each point in space. Its not hard to construct interesting functions B():
a). A 3D black and white checkerboard with 125 blocks is formed using:
B(x, y, z) = ((int)(5x) + (int)(5y) + (int)(5z)) % 2 as x, y, z vary from 0 to 1.
b). A color cube has six different colors at its vertices, with a continuously varying color at
points in between. Just use B(x, y, z) = (x, y, z) where x, y, and z vary from 0 to 1. The vertex at
(0, 0, 0) is black, that at (1, 0, 0) is red, etc.
c). All of space can be filled with such cubes stacked upon one another using B(x, y, z) =
(fract(x), fract(y), fract(z)) where fract(x) is the fractional part of the value x.
Methods for creating wood grain and turbulent marble are discussed in Chapter 14. They can be
used here as well.
In the present context we wish to paste such texture onto surfaces. To do this a bitmap is
computed using B() for each surface of the object. If the object is a cube, for instance, six
different bitmaps are computed, one for each face of the cube. Suppose a certain face of the
cube is characterized by the planar surface P + a t + b s for s, t in 0 to 1. Then use as texture
B(Px + axt + bxs, Py + ayt + bys, Pz + azt + bzs). Notice that if there is any coherence to the pattern
B() (so nearby points enjoy somewhat the same inkiness or color), then nearby points on
adjacent faces of the cube will also have nearly the same color. This makes the object truly
look like it is carved out of a single solid material.

Chapter 8

November 30, 1999

page 46

Extend Case Study 8.4 to include pasting texture like this onto the faces of a cube and an
icosahedron. Use a checkerboard texture, a color cube texture, and a wood grain texture (as
described in Chapter 14).
Form a sequence of images of a textured cube, where the cube moves slightly through the
material from frame to frame. The object will appear to slide through the texture in which it
is imbedded. This gives a very different effect from an object moving with its texture attached.
Experiment with such animations.

8.8.6. Case Study 8.6. Drawing Shadows.

(Level of Effort:III) Extend the program of Case Study 8.1 to produce shadows. Make one of
the objects in the scene a flat planar surface, on which is seen shadows of other objects.
Experiment with the projected faces approach. If time permits, develop as well the shadow
buffer approach.

8.8.7. Case Study 8.7. Extending SDL to Include Texturing.

(Level of Effort:III) The SDL scene description language does not yet include a means to
specify the texture that one wants applied to each face of an object. The keyword texture is
currently in SDL, but does nothing when encountered in a file. Do a careful study of the code in
the Scene and Shape classes, available on the books internet site, and design an approach
that permits a syntax such as
texture giraffe.bmp p1 p2 p3 p4
to create a texture from a stored image (here giraffe.bmp) and paste it onto certain faces of
subsequently defined objects. Determine how many parameters texture should require, and
how they should be used. Extend drawOpenGL() for two or three shapes so that it properly
pastes such texture onto the objects in question.

8.9. For Further Reading

Jim Blinns two JIM BLINNS CORNER books: A TRIP DOWN THE GRAPHICS PIPELINE [blinn96]
and DIRTY PIXELS [blinn98] offer several articles that lucidly explain the issues of drawing
shadows and the hyperbolic interpolation used in rendering texture. Heckberts Survey of Texture
Mapping [heckbert86] gives many interesting insights into this difficult topic. The papers Fast
Shadows and Lighting Effects Using Texture Mapping by Segal et al [segal92] and Texture
Mapping as a Fundamental Drawing Primitive by Haeberli and Segal [haeberli93] (also available
on-line: http://www.sgi.com/grafica/texmap/) provide excellent background and context.

Chapter 8

November 30, 1999

page 47

Kendrick Johnson Crime Lab Report
67% (3)
Kendrick Johnson Crime Lab Report
5 pages
Computer Graphics Using Open GL Hill Book 3rd Edition
75% (4)
Computer Graphics Using Open GL Hill Book 3rd Edition
385 pages
OSI: Physical Layer
No ratings yet
OSI: Physical Layer
36 pages
Computer Network Imp Questions
100% (1)
Computer Network Imp Questions
4 pages
2 Marks Question Bank
No ratings yet
2 Marks Question Bank
7 pages
Internetworking: Samson A. School of Electrical & Computer Engineering, Hawassa Institute of Technology
100% (1)
Internetworking: Samson A. School of Electrical & Computer Engineering, Hawassa Institute of Technology
86 pages
Chapter 3 Data and Signal
No ratings yet
Chapter 3 Data and Signal
45 pages
Mobile Banking
88% (8)
Mobile Banking
72 pages
1.2.2.3 Packet Tracer Connect and Monitor IoT Devices
No ratings yet
1.2.2.3 Packet Tracer Connect and Monitor IoT Devices
7 pages
Tutorial2 Modelanswer
No ratings yet
Tutorial2 Modelanswer
5 pages
3213 Final Fall2010
No ratings yet
3213 Final Fall2010
14 pages
VTU EC EBCS CCN Module1 Raghudathesh
100% (1)
VTU EC EBCS CCN Module1 Raghudathesh
94 pages
Delay Charts
No ratings yet
Delay Charts
9 pages
Problems and Solution_2023-2024-19-9-2024
No ratings yet
Problems and Solution_2023-2024-19-9-2024
15 pages
CS6551 CN Notes
No ratings yet
CS6551 CN Notes
216 pages
Performance Analysis of RIP, EIGRP, OSPF and ISIS Routing Protocols
No ratings yet
Performance Analysis of RIP, EIGRP, OSPF and ISIS Routing Protocols
8 pages
Computer Network Lab 1
No ratings yet
Computer Network Lab 1
61 pages
Computer Network Unit 4 Transport Layer
No ratings yet
Computer Network Unit 4 Transport Layer
23 pages
Chapter 15-17 PDF
No ratings yet
Chapter 15-17 PDF
37 pages
VLSM Exercise
No ratings yet
VLSM Exercise
2 pages
Homework 1 PDF
No ratings yet
Homework 1 PDF
2 pages
3315test2 PDF
No ratings yet
3315test2 PDF
27 pages
Nor Azimah Khalid FSKM, Uitm Shah Alam
No ratings yet
Nor Azimah Khalid FSKM, Uitm Shah Alam
50 pages
Task 3-Eder - Castañeda
100% (1)
Task 3-Eder - Castañeda
12 pages
CCN Lab Manual Vtu
50% (2)
CCN Lab Manual Vtu
21 pages
Kcs653 Mohammad Shahnawaz 1873713027
No ratings yet
Kcs653 Mohammad Shahnawaz 1873713027
121 pages
Questions Sample
100% (3)
Questions Sample
14 pages
8.2 Data Communication, Computer Network and Internet Basics
No ratings yet
8.2 Data Communication, Computer Network and Internet Basics
58 pages
Government College Women University Sialkot: Course Outline: CMP-301
No ratings yet
Government College Women University Sialkot: Course Outline: CMP-301
7 pages
Vtunotesbysri: Module 1: Application Layer
No ratings yet
Vtunotesbysri: Module 1: Application Layer
30 pages
DCN 2 Mark
100% (1)
DCN 2 Mark
7 pages
7.2.8 Packet Tracer - Verify IPv4 and IPv6 Addressing
No ratings yet
7.2.8 Packet Tracer - Verify IPv4 and IPv6 Addressing
3 pages
Chapter 2: Application Layer
100% (1)
Chapter 2: Application Layer
19 pages
Python Code Snake and Ladder1
100% (1)
Python Code Snake and Ladder1
8 pages
Chapter Two (Part 2) Data Link Layer Protocols (Q & A)
No ratings yet
Chapter Two (Part 2) Data Link Layer Protocols (Q & A)
14 pages
MMC Unit 8
No ratings yet
MMC Unit 8
59 pages
HW1
No ratings yet
HW1
9 pages
Digital Signal Processing Lab Manual
No ratings yet
Digital Signal Processing Lab Manual
171 pages
Ankon Gopal Banik: According To The Syllabus of Department of CSE, Gono Bishwabidiyalay
No ratings yet
Ankon Gopal Banik: According To The Syllabus of Department of CSE, Gono Bishwabidiyalay
36 pages
6.2 Elements of Transport Protocols PDF
No ratings yet
6.2 Elements of Transport Protocols PDF
12 pages
2marks & 16 Marks
No ratings yet
2marks & 16 Marks
4 pages
Lab Manual - Network Design Case Study
No ratings yet
Lab Manual - Network Design Case Study
143 pages
Cs3591 CN Unit 5 Notes
100% (1)
Cs3591 CN Unit 5 Notes
27 pages
Network Programming Lab Manual - Exp 1-5
No ratings yet
Network Programming Lab Manual - Exp 1-5
28 pages
Final Tutorial PDF
No ratings yet
Final Tutorial PDF
5 pages
CN Lab Manual 2018 19
No ratings yet
CN Lab Manual 2018 19
81 pages
Ch3 - Transmission Basics and Networking Media
No ratings yet
Ch3 - Transmission Basics and Networking Media
74 pages
DC&N Lab Manual
No ratings yet
DC&N Lab Manual
59 pages
10ec71 Notes
No ratings yet
10ec71 Notes
129 pages
Unit 2
No ratings yet
Unit 2
37 pages
Distributed Systems
No ratings yet
Distributed Systems
29 pages
Advanced Networking Tutorial Sheets
No ratings yet
Advanced Networking Tutorial Sheets
19 pages
CCN NOTS Final Unit - 3
100% (3)
CCN NOTS Final Unit - 3
39 pages
Computer Networks Lab Manual: V Semester
No ratings yet
Computer Networks Lab Manual: V Semester
53 pages
OSI Model: Functions of Physical Layer
No ratings yet
OSI Model: Functions of Physical Layer
13 pages
NS2 N Java Lab Manual
100% (1)
NS2 N Java Lab Manual
42 pages
Computer Networks Lab Manual
100% (2)
Computer Networks Lab Manual
61 pages
1.5.7 Packet Tracer - Network Representation
No ratings yet
1.5.7 Packet Tracer - Network Representation
3 pages
The IMS: IP Multimedia Concepts and Services
From Everand
The IMS: IP Multimedia Concepts and Services
Miikka Poikselkä
No ratings yet
CG Book 3rd Edition
No ratings yet
CG Book 3rd Edition
385 pages
Hill CH 1 Ed 3
No ratings yet
Hill CH 1 Ed 3
60 pages
Hill CH 1 Ed 3
No ratings yet
Hill CH 1 Ed 3
57 pages
Advanced Computer Systems (
No ratings yet
Advanced Computer Systems (
24 pages
BMS 110 Su17 Syllabus
No ratings yet
BMS 110 Su17 Syllabus
6 pages
Ex 4
No ratings yet
Ex 4
15 pages
Unit 1 Exam Review Sheet FA16
No ratings yet
Unit 1 Exam Review Sheet FA16
3 pages
HPC Assignment-2 Venkata Suhas Maringanti
No ratings yet
HPC Assignment-2 Venkata Suhas Maringanti
3 pages
Notice Clearance and Oath Taking
No ratings yet
Notice Clearance and Oath Taking
2 pages
Electrical and Optical Properties of Materials
No ratings yet
Electrical and Optical Properties of Materials
18 pages
Jit Production To Reduce Inventory Level-Synopsis
No ratings yet
Jit Production To Reduce Inventory Level-Synopsis
5 pages
(Childcraft - in Fifteen Volumes - 10) Field Enterprises, Inc - Art For Children (1954, Field Enterprises, Inc)
100% (1)
(Childcraft - in Fifteen Volumes - 10) Field Enterprises, Inc - Art For Children (1954, Field Enterprises, Inc)
184 pages
A European Reference Document for Langua
No ratings yet
A European Reference Document for Langua
86 pages
Green PDF
No ratings yet
Green PDF
287 pages
SDT170 User Manual en
No ratings yet
SDT170 User Manual en
141 pages
Voyager I-IOM-Cooling - Gas Fired PDF
No ratings yet
Voyager I-IOM-Cooling - Gas Fired PDF
58 pages
Single All The Way by Kate Watson
No ratings yet
Single All The Way by Kate Watson
166 pages
Creating Spaces
No ratings yet
Creating Spaces
21 pages
Stocks For The Long Run 5th All Chapter Instant Download
100% (1)
Stocks For The Long Run 5th All Chapter Instant Download
34 pages
Chapter 10 Exercise Solutions: PM N F MNMP F
No ratings yet
Chapter 10 Exercise Solutions: PM N F MNMP F
27 pages
STC-1000 Temperature Controller Operating Manual: or or or
No ratings yet
STC-1000 Temperature Controller Operating Manual: or or or
1 page
Format for Legalization of Titles from Previous Years
No ratings yet
Format for Legalization of Titles from Previous Years
4 pages
The Dragonology Handbook - Dugald A Steer
No ratings yet
The Dragonology Handbook - Dugald A Steer
88 pages
Content - E779462 - 847 Jurnal 5 Definisi Hairy Tongue PDF
No ratings yet
Content - E779462 - 847 Jurnal 5 Definisi Hairy Tongue PDF
2 pages
REPORRRRTTT123
No ratings yet
REPORRRRTTT123
5 pages
Environmental Management
No ratings yet
Environmental Management
41 pages
7. 1st Grade - Miss Laura - April 14 th to 16th
No ratings yet
7. 1st Grade - Miss Laura - April 14 th to 16th
2 pages
Transformers For Vision
No ratings yet
Transformers For Vision
28 pages
Networked Intelligence
100% (3)
Networked Intelligence
19 pages
C2 Writing CHECKLIST - RUBRICS (For Students)
No ratings yet
C2 Writing CHECKLIST - RUBRICS (For Students)
8 pages
DS - Module-1-I
No ratings yet
DS - Module-1-I
30 pages
Critical Path Method
No ratings yet
Critical Path Method
5 pages
International Master in Adult Education For Social Change: Project Coordinator
No ratings yet
International Master in Adult Education For Social Change: Project Coordinator
2 pages
Black Elegant Collage Artist Portfolio Presentation - 20240226 - 233643 - 0000
No ratings yet
Black Elegant Collage Artist Portfolio Presentation - 20240226 - 233643 - 0000
9 pages
Mathematics in The Modern World
No ratings yet
Mathematics in The Modern World
7 pages
2013 Downtown Parks Masterplan Update
No ratings yet
2013 Downtown Parks Masterplan Update
114 pages