1A) Ray tracing versus polygon scan conversion

These are the two standard methods of producing images of three-dimensional solid objects. They were covered in some detail in the Part IB course. If you want to revise them then check out FvDFH sections 14.4, 15.10 and 15.4 or F&vD sections 16.6 and 15.5. Line drawing is also used for representing three-dimensional objects in some applications. It is covered later on.

Ray tracing

<Image: a basic ray traced model showing reflection and refraction.> Ray tracing has the tremendous advantage that it can produce realistic looking images. The technique allows a wide variety of lighting effects to be implemented. It also permits a range of primitive shapes which is limited only by the ability of the programmer to write an algorithm to intersect a ray with the shape.

Ray tracing works by firing one or more rays from the eye point through each pixel. The colour assigned to a ray is the colour of the first object that it hits, determined by the object's surface properties at the ray-object intersection point and the illumination at that point. The colour of a pixel is some average of the colours of all the rays fired through it. The power of ray tracing lies in the fact that secondary rays are fired from the ray-object intersection point to determine its exact illumination (and hence colour). This spawning of secondary rays allows reflection, refraction, and shadowing to be handled with ease.

Ray tracing's big disadvantage is that it is slow. It takes minutes, or hours, to render a reasonably detailed scene. Until recently, ray tracing had never been implemented in hardware. A Cambridge company, Advanced Rendering Technologies, is trying to do just that, but they will probably still not get ray tracing speeds up to those achievable with polygon scan conversion.

Ray tracing is used where realism is vital. Example application areas are high quality architectural visualisations, and movie or television special effects.

Example

<Image: a ray traced model of a kitchen design.> This kitchen was rendered using the ray tracing program rayshade.
<Image: a close up of the kitchen sink.> These close-ups of the kitchen scene show some of the power of ray tracing. The kitchen sink reflects the wall tiles. The benchtop in front of the kitchen sink has a specular highlight on its curved front edge.
<Image: a close up of the washing machine.> The washing machine door is a perfectly curved object (impossible to achieve with polygons). The inner curve is part of a cone, the outer curve is a a cylinder. You can see the floor tiles reflected in the door. Both the washing machine door and the sink basin were made using CSG techniques (see Part 4C).
<Image: a close up of the stove and grill.> The grill on the stove casts interesting shadows (there are two lights in the scene). This sort of thing is much easier to do with ray tracing than with polygon scan conversion.

Polygon scan conversion

<Image: a scan converted model of a city - courtesy of Jon Sewell.> This term encompasses a range of algorithms where polygons are rendered, normally one at a time, into a frame buffer. The term scan comes from the fact that an image on a CRT is made up of scan lines. Examples of polygon scan conversion algorithms are the painter's algorithm, the z-buffer, and the A-buffer (FvDFH chapter 15 or F&vD chapter 15). In this course we will generally assume that polygon scan conversion (PSC) refers to the z-buffer algorithm or one of its derivatives.

The advantage of polygon scan conversion is that it is fast. Polygon scan conversion algorithms are used in computer games, flight simulators, and other applications where interactivity is important. To give a human the illusion that they are interacting with a 3D model in real time, you need to present the human with animation running at 10 frames per second or faster. Polygon scan conversion can do this. The fastest hardware implementations of PSC algorithms can now process millions of polygons per second.

One problem with polygon scan conversion is that it can only support simplistic lighting models, so images do not necessarily look realistic. For example: transparency can be supported, but refraction requires the use of an advanced and time-consuming technique called "refraction mapping"; reflections can be supported -- at the expense of duplicating all of the polygons on the "other side" of the reflecting surface; shadows can be produced, by a more complicated method than ray tracing. The other limitation is that it only has a single primitive: the polygon, which means that everything is made up of flat surfaces. This is especially unrealistic when modelling natural objects such as humans or animals. An image generated using a polygon scan conversion algorithm, even one which makes heavy use of texture mapping, will tend to look computer generated.

Examples

<Image: an SGI O2 drawn without any texture maps> <Image: an SGI O2 drawn with texture maps>
Texture mapping is a simple way of making a PSC (or a RT) scene look better without introducing lots of polygons. The image above left shows a scene without any texture maps. The equivalent scene with texture maps is shown above right. Obviously this scene was designed to be viewed with the texture maps turned on. This example shows that texture mapping can make very simple geometry look interesting to a human observer.

<Image: some floating objects in a simulated environment.> <Image: a close up of the red ball.> The image at left was generated using PSC. Texture mapping has been used to make the back and side walls more interesting. All the objects are reflected in the floor. This reflection is achieved by duplicating all of the geometry, upside-down, under the floor, and making the floor partially transparent. The close-up at right shows the reflection of the red ball, along with a circular "shadow" of the ball. This shadow is, in fact, a polygonal approximation to a circle drawn on the floor polygon and bears no relationship to the lights whatsoever.

<Image: an example of environment mapping.> <Image: an example of environment mapping.> Environment mapping is another clever idea which makes PSC images look more realistic. In environment mapping we have a texture map of the environment which can be thought of as wrapping completely around the entire scene (you could think of it as six textures on the six inside faces of a big box). The environment map itself is not drawn, but if any polygon is reflective then the normal to the polygon is found at each pixel (this normal is needed for Gouraud shading anyway) and from this the appropriate point (and therefore colour) on the environment map can be located. This example shows a silvered SGI O2 computer reflecting an environment map of the interior of a cafe.
<Image: a shot from a tank game.>
PSC is, of course, widely used in interactive games. Here we see an incautious opponent about to drive into our player's sights. The graphics are not particularly sophisticated: there are very few polygons in the scene, but the scene is made more interesting, visually, by using texture mapping. When playing the game people tend to worry more about winning (or, in some cases, not losing too badly) than about the quality of the graphics. Graphical quality is arguably more useful in selling the game to the player than in actual game play.

Line drawing

An alternative to the above methods is to draw the 3D model as a wire frame outline. This is obviously unrealistic, but is useful in particular applications. The wire frame outline can be either see through or hidden lines can be removed (FvDFH section 15.3 or F&vD section 14.2.6). In general, the lines that are drawn will be the edges of the polygons which would be drawn by a PSC algorithm.

Line drawing is generally faster than PSC unless the PSC is being done by specialist hardware. Line drawing of 3D models is used in Computer Aided Design (CAD) and in 3D model design. The software which people use to design 3D models for ray tracing tends to use both LD and PSC in its user interface. It is interesting to note that, when R&A was first written (1976), the authors had only line drawing algorithms with which to illustrate their 3D models. The only figure in the entire book which does not use exclusively line drawing is Fig. 6-52, which has screen shots of a prototype PSC system.

Out in the real world...

3D graphics finds applications in three main areas:

Visualisation generally does not require realistic looking images. In science we are usually visualising complex three dimensional structures, such as protein molecules, which have no "realistic" visual analogue. In medicine we generally prefer an image that helps in diagnosis over one which looks beautiful. PSC is therefore normally used in visualisation (although some data require voxel rendering -- see Part 4D).

Simulation uses PSC because it can generate images at interactive speeds. At the high (and very expensive) end a great deal of computer power is used. In the most expensive flight simulators (those with full hydraulic suspension and other fancy stuff) the graphics kit can cost £1,000,000 out of a total cost of about ten times that figure.

3D games (for example Quake, Unreal, Descent) use PSC because it gives interactive speeds. A lot of other "3D" games (for example SimCity, Civilisation, Diablo) use pre-drawn sprites (small images) which they simply copy to the appropriate position on the screen. This essentially reduces the problem to an image compositing operation, requiring much less processor time. The sprites can be hand drawn by an artist or created in a 3D modelling package and rendered to sprites in the company's design office. Donkey Kong Country, for example, uses sprites which were ray traced from 3D models.

You may have noticed that the previous sentence is the first mention of ray tracing in this section. It transpires that the principal use of ray tracing, in the commercial world, is in producing special effects for movies and television.

The first movie to use 3D computer graphics was Star Wars [1977]. You may remember that there were some line drawn computer graphics toward the end of the movie. All of the spaceship shots, and all of the other fancy effects, were done using models, mattes (hand-painted backdrops), and hand-painting on the actual film. Computer graphics technology has progressed incredibly since then. The recent re-release of the Star Wars trilogy included a number of computer graphic enhancements, all of which were composited into the original movie.

A recent example of computer graphics in a movie is the (rather bloodythirsty) Starship Troopers [1997]. Most of the giant insects in the movie are completely computer generated. The spaceships are a combination of computer graphic models and real models. The largest of these real models was 18' (6m) long: so it is obviously still worthwhile spending a lot of time and enrgy on the real thing.

Special effects are not necessarily computer generated. Compare King Kong [1933] with Godzilla [1998]. The plots have not changed that much, but the special effects have improved enormously: changing from hand animation (and a man in a monkey suit) to swish computer generated imagery. Not every special effect you see in a modern movie is computer generated. In Starship Troopers, for example, the explosions are real. They were set off by a pyrotechnics expert against a dark background (probably the night sky), filmed, and later composited into the movie. In Titanic [1997] the scenes with actors in the water were shot in the warm Gulf of Mexico. In order that they look as if they were shot in the freezing North Atlantic, cold breaths had to be composited in later. These were filmed in a cold room over the course of one day by a special effects studio.Film makers obviously need to balance quality, ease of production, and cost. They will use whatever technology gives them the best trade off. This is increasingly computer graphics, but computer graphics is still not useful for everything by quite a long way.

A recent development is the completely computer generated movie. Toy Story [1995] was the world's first feature length computer generated movie. Two more were released last year (A Bug's Life [1998] and Antz [1998]). Toy Story 2 [1999] will be released in the UK on 4 February 2000.

PSC or RT for SFX?

While RT gives better results than PSC, we can often get acceptable results with PSC through the use of techniques such as environment mapping and the use of lots and lots and lots of tiny polygons. The special effects industry still dithers over whether to jump in and use RT. Many special effects are done using PSC, with maybe a bit of RT for special things (giving a hybrid RT/PSC algorithm). Toy Story, for example, used Pixar's proprietary PSC algorithm. It still took between 1 and 3 hours to render each frame (although you must remember that these frames have a resolution of about 2,000 by 1,000 pixels). More expensive algorithms can be used in less time if you are rendering for television (I estimate that about one sixth of the pixels are needed compared to a movie) or if you are only rendering a small part of a big image for compositing into live action.

Finally, at SIGGRAPH 98 I had the chance to hear about the software that some real special effects companies are using:

BlueSky|ViFX
Everything is ray traced using CGI-Studio.
Digital Domain
Use ray tracing in commercial software, except when the commercial software cannot do what they want. Used MentalRay on Fifth Element [1997]; used Alias models (NURBS) passed to Lightwave (polygons) for one advertisement; used MentalRay plus Renderman for another advertisement.
Rhythm + Hues
Use a proprietry renderer, which is about 10 years old. It has been rewritten many times. They make only limited use of ray tracing.
Station X
Use Lightwave plus an internally developed renderer which is a hybrid between ray tracing and z-buffer.
Exercises
  1. Compare and contrast the capabilities and uses of ray tracing and polygon scan conversion.
  2. In what circumstances is line drawing more useful than either ray tracing or polygon scan conversion.
  3. (a) When is realism critical? Give 5 examples of applications where different levels of visual realism are necessary. Choose ray tracing or polygon scan conversion for each application and explain why you made your choice. (b) How would you determine what level of visual realism is `necessary' for a given application?
  4. "The quality of the special effects cannot compensate for a bad script." Discuss with reference to (recent) movies that you have seen.

1B) Polygon mesh management

Relevant mainly to PSC and LD.

Drawing polygons

In order to draw a polygon, you obviously need to know its vertices. To get the shading correct you also need to know its orientation: which side is the front of the polygon and which is the back. Many systems assume one-sided polygons: the front side is shaded and the back side either is coloured matt grey or black or is not even considered. This is sensible if the polygon is part of a closed polyhedron. In many applications, all objects consist of closed polyhedra; but you cannot guarantee that this will always be the case, which means that you will get unexpected results when the back sides of polygons are actually visible on screen.

The orientation of a polygon is specified by its normal vector. This does not need to be specified independently of the polygon's vertices because it can be calculated from the vertices. As an example: assume a polygon has three vertices, A, B and C. The normal vector can be calculated as: n = (C-B) x (A-B).

Any three adjacent vertices in a polygon can be used to calculate the normal vector but the order in which the vertices are specified is important: it changes whether the vector points up or down relative to the polygon. In a right-handed co-ordinate system the three vertices must be specified anti-clockwise round the polygon as you look down the desired normal vector (i.e. as you look at the front side of the polygon). Note that, if there are more than three vertices in the polygon, they must all lie in the same plane, otherwise the shape will not be a polygon!

Thus, for drawing purposes, we need to know only the vertices and surface properties of the polygon. The vertices naturally give us both edge and orientation information. The surface properties are such things as the specular and diffuse colours, and details of any texture mapping which may be applied to the polygon.

Interaction with polygon mesh data

The above is fine for drawing but, if you wish to manipulate the polygon mesh (for example, in a 3D modelling package), then is is useful to know quite a lot more about the connectivity of the mesh. For example: if you want to move a vertex, which is shared by four polygons, you do not want to have to search through all the polygons in your data structure trying to find the ones which contain a vertex which matches your vertex data.

The winged-edge data structure is particularly useful for handling polygon mesh data. It contains explicit links for all of the relationships between vertices, edges and polygons, thus making it easy to find, for example, which polygons are attached to a given vertex, or which polygons are adjacent to a given polygon (by traversing the edge list for the given polygon, and finding which polygon lies on the other side of each edge).

The vertex object contains the vertex's co-ordinates, a pointer to a list of all edges of which this vertex is an end-point, and a pointer to a list of all polygons of which the vertex is a vertex.

The polygon object contains (a pointer to) the polygon's surface property information, a pointer to a list of all edges which bound this polygon, and a pointer to an ordered list of all vertices of the polygon.

The edge object contains pointers to its start and end vertices, and pointers to the polygons which lie to the left and right of it.

<Image: diagram showing the winged edge data structure as presented in the lecture>

The above diagram shows just one possible implementation of a polygon mesh data structure. FvDFH section 12.5.2 describes another winged-edge data structure which contains slightly less information, and therefore requires more accesses to find certain pieces of information than the one shown above. The implementation that would be chosen depends on the needs of the particular application which is using the data structure. Another alternative would be to attach shading information to vertices rather than to polygons. This could then be used to Gouraud shade or Phong shade the polygons.

F&vD section 13.2 also contains a bit of information on polygon meshes.

Hardware PSC quirks

A piece of PSC hardware, such as the Silicon Graphics Reality Engine, will generally consist of a geometry engine and a rendering engine. The geometry engine will handle the transformations of all vertices and normals (and possibly handle some of the shading calculations). The rendering engine will implement the PSC algorithm on the transformed data.

Triangles only

When making a piece of hardware to render a polygon, it is much easier to make the hardware handle a fixed number of vertices per polygon, than a variable number. Many hardware implementations thus simply implement triangle drawing. This is not a serious drawback. Polygons with more vertices are simply split into triangles.

The triangle strip set and triangle fan set

In addition to simple triangle drawing, Silicon Graphics machines implement both the triangle strip set and triangle fan set to speed up processing through the geometry engine. Each triangle in the set has two vertices in common with the previous triangle. Each vertex is transformed only once by the geometry engine, giving a factor of three speed up in geometry processing.

For example, assume we have triangles ABC, BCD, CDE and DEF. In standard triangle rendering, the vertices would be sent to the geometry engine in the order ABC BCD CDE DEF; each triangle's vertices being sent separately. With a triangle strip set the vertices are sent as ABCDEF; the adjacent triangles' vertices overlapping.

A triangle fan set is similar. In the four triangle case we would have triangles ABC, ACD, ADE and AEF. The vertices would again be sent just as ABCDEF. It is obviously important that the rastering engine be told whether it is drawing standard triangles or a triangle strip set or a triangle fan set.

Exercises
  1. Calculate both surface normal vectors (left-handed and right-handed) for a triangle with points (1, 1, 0), (2, 0, 1), (-1, -2, -1).
  2. Confirm that the following statements are correct for a reasonable polygon mesh:
    1. A vertex belongs to at least two edges.
    2. A vertex is a vertex of at least one polygon.
    3. An edge has exactly two end points.
    4. An edge is an edge of either one or two polygons.
    5. A polygon has at least three vertices.
    6. A polygon has at least three edges.
  3. Work out the algorithm that is required to modify a winged-edge data structure when an edge is split. You may ignore surface property information for the polygons and you may assume that the edge that is split is split exactly in half. The algorithm could by called by the function call:
    split_edge( vertex_list v, edge_list e, polygon_list p, edge edge_to_split)
    where the winged-edge data structure is made up of the three linked lists of objects (vertices, edges, and polygons).

2A) Ray tracing primitives

Relevant mainly to RT.

A primitive is a shape for which a ray-shape intersection routine has been written. More complex objects can be built out of the primitives. Most ray tracers will have a variety of primitives. They are limited only by the ability of the programmer to write a function to analytically intersect a ray with the surface of the shape.

Common primitives

Plane
The infinite plane is a simple object with which to intersect a ray (see Part IB notes for details). On its own it can represent boundary objects such as the ground or the sky or perhaps an infinite wall. Intersection with the infinite plane is a useful building block in a ray tracing system.

<Image: ray traced polygon> Polygon
Having the polygon as a ray tracing primitive allows a ray tracer to render anything that a PSC algorithm could. To find the intersection of a ray with a polygon, first find the intersection of the ray with the infinite plane in which the polygon lies. Then ascertain whether the intersection lies inside or outside the polygon: this is a reasonably straightforward two dimensional graphics operation.

<Image: ray traced box> Box
A box is essentially six polygons and could be ray traced as such. However, intersection with an axis-aligned box can be optimised. Any box can be axis-aligned by appropriate transformations. We can thus write a routine to intersect an arbitrary ray with an axis-aligned box and then transform the ray under consideration in exactly the same way as we transform the box which we are trying to intersect with it. This sort of idea generalises neatly to the concept of specifying any object in a convenient object coordinate system and then applying transforms to the whole object to place it at the appropriate point in the world coordinate system.

<Image: ray traced sphere> Sphere
The sphere is the simplest finite object with which to intersect a ray. Practically any ray tracing program will include the sphere as a primitive. Scaling a sphere by different amounts along the different axes will produce an ellipsoid: a squashed or stretched sphere. There is thus no need to include the ellipsoid as a primitive provided that your ray tracer contains the usual complement of transformations. (It would be a poor ray tracer if it did not!)

<Image: ray traced cylinder> Cylinder
Intersecting a ray with an infinitely long cylinder is practically as easy as intersecting one with a sphere. The tricky bit, if it can be called that, is to intersect a ray with a more useful finite length cylinder. This is achieved by intersecting the ray with the appropriate infinitely long cylinder and then ascertaining where along the cylinder the intersection lies. If it lies in the finite length in which you are interested then keep the intersection. If it does not then ignore the intersection. Note that the ray tracer used to render the accompanying image has cylinders without end caps. This is the correct result if you follow the procedure outlined in this paragraph. Adding end caps to your cylinders requires extra calculations.

<Image: ray traced cone> Cone
Cones are very like cylinders. Like the infinite cylinder, there is a simple mathematical definition of an infinite cone which makes it easy to write a ray-cone intersection algorithm. Note that a cone does not need to have a point -- it can be truncated short of its `top', as illustrated in the accompanying image. The particular ray tracer used does not add end caps to cones.

<Image: ray traced disc> Disc
The disc is not a common ray tracing primitive, but is necessary in ray tracers which implement cones and cylinders without end caps. It is handled in much the same way as the polygon. The routine to check if the intersection with the plane lies inside or outside of the disc is simpler than the equivalent routine for the polygon. In the accompanying ray traced image, the disc has been positioned so that it just catches the light -- this illustrates how specular reflection varies across a completely flat surface. The sphere, cone and torus images illustrate how specular reflection varies across three curved surfaces.

<Image: ray traced torus> Torus
Toroids are reasonably rare in real life (doughnuts and tyre inner tubes notwithstanding). They are somehow alluring to the kinds of people who implement ray tracers and, having a reasonably straightforward mathematical definition, are reasonably simple to implement. They thus appear as primitives in many ray tracers. They become more useful when combined with Constructive Solid Geometry (see Part 4C).

Ray-primitive intersections

SMAG section 1 contains the mathematics of ray-primitive intersections and details of how you would go about converting a primitive object into polygons for PSC.

Exercises
  1. Work out mathematical equations to define a plane, a sphere, an infinitely long cylinder, an infinitely long cone, and a torus. You will find it helpful to centre each primitive at the origin and to align it in a sensible way with respect to the coordinate axes. You may like to use cylindrical polar coordinates where these prove useful. As a starting point remember that a circle can be represented by the equation:x^2+y^2=r^2
  2. Work out a way to intersect a ray with each of the five primitives. You may assume that you are provided with functions to find the roots of linear, quadratic, cubic and quartic equations.
  3. Work out if there exists a faster intersection algorithm for an axis aligned 2x2x2 unit box than just six separate polygon intersection calculations (N.B. I don't know whether there is or not).
  4. [1999/7/11] (a) Give a parametric definition of a torus centred at the origin and aligned with the coordinate axes. (b) Outline how you would find the first intersection point, if any, of a ray with the torus from the previous part.
  5. Show how to convert a cylinder into a polygon mesh. What changes do you have to make if the mesh may contain only triangles?
  6. Show how to convert a torus into a polygon mesh.
  7. Show how to convert a sphere into a triangle mesh. How can you get the most even distributiuon of triangle vertices across the sphere?

2B) Conics, quadrics, and superquadrics

The ray tracing primitives, described in Part 2A, have relatively simple mathematical definitions. This is what makes them attractive: the simple mathematical definition allows for simple ray-object intersection code. Following from this, it would seem logical to investigate other shapes with simple mathematical definitions. Spheres, cones and cylinders are part of a more general family of parametric surfaces called quadrics (N.B. tori are not quadrics). Quadrics are the 3D analogue of 2D conics. We describe these general families below, but it turns out that they are of little practical use. It would seem that the general quadrics are a "dead end" in graphics research.

Conics

A conic is a two dimensional curve desribed by the general equation:
Ax^2+Bxy+Cy^2+Dx+Ey+F=0
This general form can be rotated, scaled, and translated so that it is aligned along the axes of the coordinate system. It will then have the simpler equation:
ax^2+by^2=k or ax^2+by=k
The useful conics are the ellipse (of which the circle is a special case), the hyperbola, and the parabola. For more details see R&A section 4-10, especially Table 4-8 on page 242. Table 4-8 is included in the handout.

Quadrics

The quadrics are the three dimensional analogue of the conics. The general equation is:
Ax^2+By^2+Cz^2+Dxy+Eyz+Fzy+Gx+Hy+Jz+K=0
This general form can be rotated, scaled, and translated so that it is aligned along the axes of the coordinate system. It will then have the simpler equation:
ax^2+by^2+cz^2=k or ax^2+by^2+cz=k
The useful conics are the ellipsoid (of which the sphere is a special case), the infinite cylinder, and the infinite cone. Various hyperboloids, and paraboloids are also defined by these equations, but these have little real use unless one is designing satellite dishes (paraboloid), headlamp reflectors (also paraboloid), or power station cooling towers (hyperboloid). For more details see R&A section 6-4, especially Figure 6-18 on page 403. Figure 6-18 is included in the handout.

Superquadrics

These are an extension of quadrics, where the power on the coordinate does not have to be 2. The general form of a superquadric centred at the origin and aligned along the coordinate axes is:
(ax)^n+(by)^n+(cz)^n=k

Super-ellipsoids tend to be the only members of this family that are actually used, and even they are only used in very limited areas. The effect of n on a super-ellipsoid is roughly as follows: n=2 is a standard ellipsoid; n<2 is a more pointy version, the "points" being along the main axes; n>2 becomes closer to a box as n increases; n=1 is allegedly an octahedral shape; and n<1 is truly pointy along the main axes.

The interested student may like to have a quick look at Alan Barr's two papers on superquadrics. The papers can be found in the Computer Laboratory library in IEEE Transactions on Computer Graphics and Applications volume 1, number 1 (January 1981), pages 11-23, and volume 1, number 3 (July 1981), pages 41-47.

Brian Wyvill describes a use of super-ellipsoids on pages 264 and 265 of "A Computer Animation Tutorial" in Computer Graphics Techniques: Theory and Practice, Rogers and Earnshaw (editors), Springer-Verlag, 1990, ISBN 0-387-97237-4. This extract is included in the handout.