parts and the shading of the objects.

Until now, most

effort has been spent in the search for fast hidden surface
removal algorithms. With the development of these
algorithms, the programs that produce pictures are
becoming remarkably fast, and we may now turn to the
Graphics and W. Newman
search for algorithms to enhance the quality of these
Image Processing Editor
pictures.
Illumination for In trying to improve the quality of the synthetic
images, we do not expect to be able to display the object
Computer Generated exactly as it would appear in reality, with texture, over-
cast shadows, etc. We hope only to display an image
Pictures that approximates the real object closely enough to
provide a certain degree of realism. This involves some
understanding of the fundamental properties of the
Bui Tuong Phong human visual system. Unlike a photograph of a real
University of Utah world scene, a computer generated shaded picture is
made from a numerical model, which is stored in the
computer as an objective description. When an image is
then generated from this model, the human visual sys-
tem makes the final subjective analysis. Obtaining a
close image correspondence to the eye's subjective
The quality of computer generated images of three- interpretation of the real object is then the goal. The
dimensional scenes depends on the shading technique computer system can be compared to an artist who
used to paint the objects on the cathode-ray tube screen. paints an object from its description and not from direct
The shading algorithm itself depends in part on the observation of the object. But unlike the artist, who can
method for modeling the object, which also determines correct the painting if it does not look right to him, the
the hidden surface algorithm. The various methods of computer that generates the picture does not receive
object modeling, shading, and hidden surface removal feedback about the quality of the synthetic images,
are thus strongly interconnected. Several shading tech- because the human visual system is the final receptor.
niques corresponding to different methods of object This is a subjective domain. We must at the outset
modeling and the related hidden surface algorithms are define the degree of realism we wish to attain, and fix
presented here. Human visual perception and the funda- certain goals to be accomplished. A m o n g these goals
mental laws of optics are considered in the development are:
of a shading rule that provides better quality and in- 1. "Real time" display of dynamic color pictures of
creased realism in generated images. three-dimensional objects. A real time display system
Key Words and Phrases: computer graphics, graphic is one capable of generating pictures at the rate of at
display, shading, hidden surface removal. least 30 frames a second.
CR Categories: 3.26, 3.41, 8.2 2. Representation of objects made of smooth curved
surfaces.
3. Elimination or attenuation of the effects of digital
Introduction sampling techniques.
The most important consideration in trying to attain
This .paper describes several approaches to the pro- these goals is the object modeling technique.
duction of shaded pictures of solid objects. In the past
decade, we have witnessed the development of a number
of systems for the rendering of solid objects by com- Existing Shading Techniques
puter. The two principal problems encountered in the
design of these systems are the elimination of the hidden Methods of Object Modeling
Copyright @ 1975, Association for Computing Machinery, Inc. Image quality depends directly on the effectiveness
General permission to republish, but not for profit, all or part of the shading algorithm, which in turn depends on the
of this material is granted provided that ACM's copyright notice
is given and that reference is made to the publication, to its date method of modeling the object. Two principal methods
of issue, and to the fact that reprinting privileges were granted of object description are commonly used :
by permission of the Association for Computing Machinery. 1. Surface definition using mathematical equations.
This research was supported in part by the University of Utah
Computer Science Division and the Advanced Research Projects 2. Surface approximation by planar polygonal mosaic.
Agency of the U.S. Department of Defense, monitored by the Several systems have been implemented to remove
Rome Air Development Center, Griffiss Air Force Base, NY hidden parts for mathematically defined curved surfaces
13440, under Contract F30602-70-C-0300. Author's address:
Digital Systems Laboratory, Stanford University, Stanford, CA [1, 2, 3, 4, 5]. With these systems, exact information at
94305. each point of the surface can be obtained, and the result-

311 Communications June 1975

of Volume 18
the ACM Number 6
ing computer generated pictures are most realistic. The Fig. 1. A cone represented by means of planar approximation.
class of possible surfaces is restricted, however, and the
computation time needed to remove the hidden parts
and to perform shading is very large. Up to the present
time, these systems have usually considered the class of
surfaces represented by quadric patches. Although
higher degree surfaces are desirable and are sometimes
necessary to model an object, they have not been taken
into consideration due to an increase in computation
time to remove hidden surfaces and to perform shading
computations. Even when only quadric surfaces are
considered, the implementation of a real time display
system using this type of model is too expensive and
complex.
A simple method of representing curved surfaces and
objects of arbitrary shape is to approximate the surfaces
with small planar polygons; for example, a cone might
be represented as shown in Figure 1. This type of repre-
sentation has the advantage that it avoids the problem,
posed by mathematically curved surface approaches, of
solving higher order equations.
Planar approximation also offers the only means of
reducing hidden surface computation to within reason-
able bounds, without restricting the class of surfaces
that can be represented. For this reason, all recent the polygon, one screen resolution unit in height, that
attempts to devise fast hidden surface algorithms have lies on the scan line.
been based on the use of this approximation for curved Newell, Newell, and Sancha adopt a different ap-
surfaces; these algorithms have been summarized and proach, using a frame buffer into which the object is
classified by Sutherland et al. [6]. The next section dis- painted, face by face. The hidden surface problem is
cusses their influence on the way shading is computed. solved by painting the farthest face first, and the nearest
While planar approximation greatly simplifies last. Each face is painted scan line by scan line, starting
hidden surface removal, it introduces several major at the top of the face.
problems in the generation of a realistic displayed F r o m the shading aspect, the important attribute of
image. One of these is the contour edge problem: the these algorithms is that they both generate information
outline or silhouette of a polygonally approximated scan line by scan line in order to display the faces of an
object is itself a polygon, not a smooth curve. The other object. This information is in the form of segments, one
problem is that of shading the polygons in a realistic screen resolution unit high, on which the shading com-
manner. This paper is concerned with the shading putation may then be performed. The main differences
problem; the contour edge problem is discussed by the between the algorithms, from the point of view of
author and F.C. Crow in [7]. shading, are (a) the order in which the segments are
generated, and (b) the fact that Watkins generates each
Influence of Hidden Surface Algorithms screen dot only once, whereas the NewelI-Sancha al-
The order in which a hidden surface algorithm com- gorithm may overwrite the same dot several times.
putes visible information has a decided influence on the
way shading is performed. For example Warnock, who Shading with the Polyhedral Model
developed one of the first such algorithms [8], com- When planar polygons are used to model an object,
puted display data by a binary subdivision process: this it is customary to shade the object by using the normal
meant that the order of generating display data was vectors to the polygons. The shading of each point on a
largely independent both of the order of scanning the polygon is then the product of a shading coefficient for
display and of the order of the polygons in memory. the polygon and the cosine of the angle between the
This made it difficult to perform effective shading on polygon normal and the direction of incident light. This
curved objects. cosine relationship is known in optics as the "cosine
The two major advances in the development of fast law," and allows us to compute the shading Sp for a
hidden surface algorithms have been made by Watkins polygon p as
[9] and by Newell, Newell, and Sancha [10]. Watkins
sp = Cpcos(i), (1)
generates the displayed picture scan line by scan line.
On each scan line he computes which polygons intersect where Cp is the reflection coefficient of the material o f p
the scan line, and then computes the visible segment of relative to the incident wavelength, and i is the incident
each polygon, where this segment is the visible strip of angle.

312 Communications June 1975

of Volume 18
the ACM Number 6
Fig. 2. An example of the use of Newell, Newell, and Sancha's This shading offers only a very rough approximation
shading technique, showing transparency and highlight effects.
of the true physical effect. It does not allow for any of
the specular properties of the material, i.e. the ability of
the material to generate highlights by reflection from its
outer surface, and the position of the observer, which is
ignored. A more serious drawback to this method, how-
ever, is the poor effect when using it to display smooth
curved surfaces. The cosine law rule is appropriate for
objects that are properly modeled with planar surfaces,
such as boxes, buildings, etc., but it is inappropriate for
smoothly curved surfaces such as automobile bodies.
This does not mean, however, that we should abandon
the use of such a polygon-oriented shading rule and
search for a different rule for curved surfaces. Recent
research in shading techniques demonstrates that signifi-
cant results can be achieved by using the basic shading
rule of eq. (1) and modifying the results to reduce the
discontinuities in shading between adjacent polygons.

1. Warnock's shading. As three-dimensional objects

are projected onto the cathode-ray tube screen, the
depth sensation is lost, and the images of those objects
appear flat. In order to restore the depth sensation, two
Fig. 3. Computation of the shading at point R using the Gouraud effects were simulated by Warnock:
method. There are two successive linear interpolations: (1) across I. Decreasing intensity of the reflected light from the
polygon edges, i.e. P between A and B, Q between A and D; and (2)
along the scan line, i.e. R between P and Q. object with the distance between the light source and the
object.

2. Highlights created by specular reflection.

Warnock placed the light source and the eye at the
same position, so that the shading function was the sum
of two terms, one for the normal "cosine" law, and the
other term for the specularly reflected light. The result-
ing pictures have several desirable attributes; for exam-
ple, identical parallel faces, located differently in space,
will be shaded at different intensities, and facets which
face directly toward the light source are brighter than
Fig. 4. Gouraud shading, applied to approximated cone of Fig. 1. adjacent facets facing slightly away from the incident
light. However, the polygonal model gives a discontinu-
ity in shading between faces of an approximated curved
surface. When a curved surface is displayed, the smooth-
ness of the curved surface is destroyed by this discon-
tinuity. This is clearly visible in Figure 1.

2. Newell, Newell, and Sancha's shading. Newell,

Newell, and Sancha presented some ideas on creating
transparency and highlights. From observations in the
real world, they found that highlights are created not
only by the incident light source but also by the reflec-
tion of light from other objects in the scene; this is
especially true in the case of objects made of highly
reflective or transparent materials. In the Newell-
Sancha model, curved surfaces are approximated with
planar polygons. Unfortunately, the ability to generate
highlights is severely limited due to the inability to vary
light intensity over the surface of any single polygon.
This problem is apparent in Figure 2.

3. Gouraud's shading. While working on a technique

to represent curved objects made of "Coons surfaces"

313 Communications June 1975

of Volume 18
the ACM Number 6
 or "Bezier patches," Gouraud [11] developed an al- Fig. 5. Normal at a point Fig. 6. Shading at a point.
gorithm to shade curved surfaces. With his algorithm, along an edge.
,¢
a surface represented by a patch is approximated by ILIGHT
REFLECTED
RAY /
/ EYE
SOURCE Np /
polygonal planar facets. Gouraud computes information /
\ /
about the curvature of the surface at each vertex of each \ /
^ ^ ^ ~ X,
of these facets. F r o m the curvature, a shade intensity is A N, ^ \ ~ r
*, Ni/4 /2 N~/4 ~' INCIDENT
computed and retained. For example, the shade intensity "0 . . . . _W~N_++___~_N... . . . . . . "I RAY ~ .

may be computed for each vertex using eq. (1), with i as

P
the angle between the incident light and the normal to Po
the surface at this vertex. When the surface is displayed,
this shade intensity is linearly interpolated along the
edge between adjacent pairs of vertices of the object.
The shade at a point on the surface is also a linear inter- Unfortunately, because of visual perception effects,
polation of the shade along a scan line between inter- the reduction of polygon size is not as beneficial as
sections of the edges with a plane passing through the might be expected, The particular effect responsible is
scan line (Figure 3). This very simple method gives a the Mach Band effect. Mach established the following
continuous gradation of shade over the entire surface, principle:
which in most cases restores the smooth appearance. An
Wherever the light-intensity curve of an illuminated surface (the
example of Gouraud's shading is shown in Figure 4. light intensity of which varies in only one direction) has a concave
With the introduction of the Gouraud smooth shading or convex flection with respect to the axis of the abscissa, that
technique, the quality of computer-generated images particular place appears brighter or darker, respectively, than its
surroundings [E. Mach, 1865].
improved sufficiently to allow representation of a large
variety of objects with great realism. Problems still Whenever the slope of the light intensity curve changes,
exist, however, one of which is the apparent discon- this effect appears. The extent to which it is noticeable
tinuity across polygon edges. On surfaces with a high depends upon the magnitude of the curvature change,
component of specular reflection, highlights are often but the effect itself is always present.
inappropriately shaped, since they depend upon the Without the Mach Band effect, one might hope to
disposition and shape of the polygons used to approxi- achieve accurate shading by reducing the size of poly-
mate a curved surface and not upon the curvature of.the gons. Unfortunately the eye enhances the discontinuities
object surface itself. The shading of a surface in motion over polygon edges, creating undesired areas of appar-
(in a computer generated film) has annoying frame to ent brightness along the edges. Therefore unless the size
frame discontinuities due to the changing orientation of of the displayed facets is shrunk to a resolution point,
the polygons describing the surface. Also the shading increasing the number of facets does not solve the
algorithms are not invariant under rotation. problem. Using the Gouraud method to interpolate the
Frame-to-frame discontinuities of shade in a com- shade linearly between vertices, the discontinuities of
puter generated film are illustrated in the following the shading function disappear, but the Mach Band
situation. A curved surface is approximated with planar effect is visible where the slope of the shading function
facets. When this surface is in motion, all the facets changes. This can be seen in Figure 4. The subjective
which are perpendicular to the direction of the light take discontinuity of shade at the edges due to the Mach
on a uniform shade. In the next frame the motion of the Band effect then destroys the smooth appearance of the
object brings these facets into a different orientation curved surface.
toward t h e light, and the intensity of the shade across A better shading rule is therefore proposed for dis-
their surfaces varies continuously from one end to the playing curved surfaces described by planar polygons.
other. Thus the surface appears to change from one with This new technique requires the computation of the
highlights to one of uniform shade. Moreover, the normal to the displayed surface at each point. It is
position of these highlights is not steady from frame to therefore more expensive in computation than
frame as the object rotates. Gouraud's technique; but the quality of the resulting
picture, and the accuracy of the displayed highlights, is
Mach Band Effect much improved.
Many of the shading problems associated with
planar approximation of curved surfaces are the result Using a Physical Model
of the discontinuities at polygon boundaries. One might
expect that these problems could be avoided by reducing Specular Reflection
the size of the polygons. This would be undesirable, of If the goal in shading a computer-synthesized image
course, since it would increase the number of polygons is to simulate a real physical object, then the shading
and hence would increase both the memory require- model should in some way imitate real physical shading
ments for storing the model and the time for hidden situations. Clearly the model of eq. (1) does not ac-
surface removal. complish this. As mentioned before, it completely

314 Communications June 1975

of Volume 18
the ACM Number 6
Fig. 7(a). Determination of the reflected light. surface of a polygon is achieved in the same way as the
,Y
computation of the shading at that point with the
Gouraud technique. The normal to the visible surface at
a point located between two edges is the linear inter-
VR
polation of the normals at the intersections of these two
edges with a scan plane passing through the point under
consideration. Note that the general surface normal is
quadratically related to the vertex normal.
I-Z
\ q-- :/ - F r o m the approximated normal at a point, a shading
\ I"-,,< i i ..... function determines the shading value at that point.

The Shading Function Model

Fig. 7(b). Projections of the reflecte~ light. In computer graphics, a shading function is defined
as a function which yields the intensity value of each
point on the body of an object from the characteristics
of the light source, the object, and the position of the
observer.
Taking into consideration that the light received by
[ . . . . .
the eye is provided one part by the diffuse reflection and
I ........... one part by the specular reflection of the incident light,
the shading at point P (Figure 6) on an object can be
LIGHT
I
computed as:
X • LZ
xR XN
Sp = Cp[cos(i) (1 - d) +d] q- W(i) [cos(s)] ", (3)
ignores both the position of the observer and the specu-
lar properties of the object. Even with the improve- where:
ments introduced by Gouraud, which provide remark- Cp is the reflection coefficient of the object at point P
ably better shading, these properties are still ignored. for a certain wavelength.
The first step in accounting for the specular proper- i is the incident angle.
ties of objects and the position of the observer is to d is the environmental diffuse reflection coefficient.
determine the normal to the surface at each point to be W(i) is a function which gives the ratio of the specular
shaded, i.e. at each point where a picture element of the reflected light and the incident light as a function
raster display projects onto the surface. It is only with of the incident angle i.
this knowledge that information about the direction of s is the angle between the direction of the reflected
reflected rays can be acquired, and only with this in- light and the line of sight.
formation can we model the specular properties of n is a power which models the specular reflected
objects. It is evident from the preceding discussion, light for each material.
however, that our polyhedral model provides informa- The function W(i) and the power n express the
tion about normals only at the vertices of polygons. specular reflection characteristics of a material. For a
Thus the first step in improving our shading model is to highly reflective material, the values of both W(i) and n
devise a way to obtain the normal to the surface for each are large. The range of W(i) is between 10 and 80
raster unit. percent, and n varies from 1 to 10. These numbers are
empirically adjusted for the picture, and no physical
Computation of the Normal at a Point on the Surface justifications are made. In order to simplify the model,
The normal at each vertex can be approximated by and thereby the computation of the terms cos(i) and
either one of the methods described by Gouraud [I0]. cos(s) of formula (3), it is assumed that:
It is now necessary to define the normal to the surface 1. The light source is located at infinity; that is, the
along the edges and at a point on the surface of a poly- light rays are parallel.
gon. 2. The eye is also removed to infinity.
The normal to the surface at a point along the edge With these two considerations, the values of cos(i)
of a polygonal model is the result of a linear interpola- and cos(s) of the shading function in (3) can be re-
tion to the normals at the two vertices of that edge. An written as: cos(i) = kNp / IN~I and cos(s) = uRn, / iRpl
example is given in Figure 5: the normal Nt to the where k and u are respectively the unit vectors in the
surface at a point between the two vertices P0 and P1 is direction of the light and the line of sight, Np is the
computed as follows: normal vector at P, and Rp is the reflected light vector at
P.
Nt = tN1 q- (1--t)N0, (2)
The quantity k N p / [NpJ can be referred to as the
where t = 0 at No a n d t = 1 a t N x . projection of a normalized vector N~ on an axis parallel
The determination of the normal at a point on the to the direction of the light. If INp[ is unity, the previous
315 Communications June 1975
of Volume 18
the ACM Number 6
Fig. 8. Improved shading, applied to approximated cone of Fig. 1. Fig. 9. Improved shading, applied to the example of Figure 2.

quantity is one component of the vector Np in a co-

ordinate system where the direction of light is parallel By simple trigonometry, we obtain the following
to one axis. In this case, the quantity u R p / IRp[ can be expressions:
obtained directly from the vector N~ in the following Z~ = cos(2i) = 2[cos(i)] 2 -- 1 = 2Z, 2 - 1, (6)
way.
Let us consider a Cartesian coordinate system having Xr2q - Yr ~ = [sin(2i)]2= 1 - [cos(2i)] 2. (7)
the origin located at point P and having the z axis
F r o m (4) and (7), we obtain:
parallel to the light but opposite in direction (Fig. 7(a)).
We have the following assumptions about the X~ = 2ZnXn , Yr = 2Z, Y , , 0 < Zn <_ 1.
model:
1. The normalized vector Np makes an angle i with the The three components of Rp are then known in the
z axis, and the reflected light vector Rp makes an angle light source coordinate system. The projection of the
2i with the same axis. vector Rp onto the z-axis of the eye coordinate system
2. Only incident angles less than or equal to 90 degrees may be found by a simple dot product of the reflected
are considered in the shading computation. For a vector with this z-axis. The component of Rp on an axis
greater angle, this means that the light source is behind parallel to the line of sight is the value of the cosine of
the front surface. In the case where a view of the back the angle between the reflected light and the line of
surface is desired when it is visible, it can be assumed sight. The value of this cosine will be used in the sim-
that the normal will always point toward the light ulation of the specular reflection.
source. This method of calculating the direction of the re-
3. If k is the unit vector along the P Z axis, then by flected light for each point from the orientation of the
simple geometry, it may be shown that the three vectors normal is preferred over the computation of the re-
k, Np, and Rp are coplanar. flected light vector at vertices and the subsequent inter-
4. The two vectors Np and Rp are of unit length. polation of them in the same way as the normal. It is
F r o m assumption (3), the projections of the vectors faster and it requires less storage space than the inter-
Np and Rp onto the plane defined by (PX, P Y) are polation scheme.
merged into a line segment (Figure 7(b)). Therefore, With the described method, the shading of a point is
computed from the orientation of the approximated
Xr/Yr = X./r., (4) normal; it is not a linear interpolation of the shading
where X~, 2"., Yr, and Y. are respectively the compo- values at the vertices. Therefore, a better approximation
nents of R~ and Np in the x and y directions. of the curvature of the surface is obtained, and high-
F r o m assumptions (1) and (2), the component lights due to the simulation of specular reflection are
Z . of Np is: properly rendered. Examples of application of the
shading technique are shown in Figures 8 and 9. Figure
Z . = cos(i), (5) 10 compares a display generated by this technique with a
where 0 <_ i < 90 degrees. photograph of a real object.

316 Communications June 1975

of Volume 18
the ACM Number 6
Fig. 10(a). A sphere displayed with the improved shading. that this effect is visible whenever there is a great change
in the slope of the intensity distribution curve, even if
the curve has a continuous first derivative. When a
higher degree interpolation curve is used, it will make
the presence of the edges unnoticeable, although it will
still give some Mach Band effect.
When a comparison was made of pictures of the
same object generated with different shading tech-
niques, it was found that little difference existed between
pictures generated with the new shading and the ones
created with a cubic interpolant curve for the shading
computation. Furthermore, as time is the critical factor
in a real time dynamic picture display system, the use of
a high degree interpolation curve does not seem to be
possible at the moment with the current techniques to
compute the coefficients of such a function.
A hardware implementation of this shading model
would of course require more hardware than the simpler
Gouraud method. The Gouraud model needs one inter-
polator for the shading function. It must compute a new
shading value for each raster unit, and hence must be
Fig. 10(b). A photograph of a real sphere. very high speed to drive a real time display. The model
proposed here requires three of these interpolators
operating in parallel. In addition, since the results of the
interpolation do not yield a unit vector, and since eqs.
(6), (7), and (8) require a unit normal vector, some extra
hardware is necessary to "normalize" the outputs of the
interpolators. This requires a very fast mechanism for
obtaining square roots. None of these problems is too
difficult to solve; and judging from the improvements in
image quality obtained using the new model, it may well
be worth the extra expense to provide such hardware in
applications for which real time display is important.
Received November 1975; revised March 1975

References
1. MAGI, Mathematical Applications Group Inc. 3-D simulated
graphics. Datamation 14 (Feb. 1968), 69.
2. Comba, P.G. A procedure of detecting intersections of three-
dimensional objects. Rep. 39,020, IBM New York Scientific
Center, Jan. 1967.
3. Weiss, R.A. BE VISION, a package of IBM 7090 FORTRAN
programs to draw orthographic views of combinations of plane
and quadric surfaces. J. A C M 13, 2 (Apr. 1966), 194-204.
4. Mahl, R. Visible surface algorithm for quadric patches.
IEEE Trans. C-21, (Jan. 1972), 1-4.
5. Catmull, E.E. A subdivision algorithm for computer display
of curved surfaces. Ph.D th., Dep. of Comput. Sci., U. of Utah.
Conclusion 6. Sutherland, I.E., Sproull, R.F., and Schumacker, R.A. A
characterization of ten-hidden surface algorithms. Computing
Surveys 6 (Mar. 1974), 1-56.
The linear interpolation scheme used here to ap-
7. Bui Tuong Phong and Crow, F.C. Improved rendition of
proximate the orientation of the normal does not polygonal models of curved surfaces. To be presented at the
guarantee a continuous first derivative of the shading joint USA-Japan Computer Conference.
function across an edge of a polygonal model. In ex- 8. Warnock, J.E. A hidden-line algorithm for halftone picture
representation. Dep. of Comput. Sci., U. of Utah, TR 4-15,
treme cases where there is an abrupt change in the 1969.
orientation of two adjacent polygons along a common 9. Watkins, G.S. A real-time visible surface algorithm. Dep. of
edge, the subjective brightness due to the Mach Band Comput. Sci., U. of Utah, UTEC-CSc-70-101, June 1970.
10. Newell, M.E., Newell, R.G., and Sancha, T.L. A new
effect will be visible along this edge. However, this approach to the shaded picture problem. Proc. ACM 1973
effect is much less visible in the described model than in Nat. Conf.
the Gouraud smooth shading model. Also, an interest- 11. Gouraud, H. Computer display of curved surfaces. Dep. of
Comput. Sci., U. of Utah, UTEC-CSc-71-113, June 1971. Also
ing fact discussed previously on Mach Band effect shows in IEEE Trans. C-20 (June 1971), 623-629.

317 Communications June 1975

of Volume 18
the ACM Number 6