CSE167: Computer Graphics
Instructor: Steve Rotenberg
UCSD, Fall 2005







Ray tracing is a powerful rendering technique that is the foundation of many modern photoreal rendering algorithms
The original ray tracing technique was proposed in 1980 by Turner
Whitted, although there were suggestions about the possibility in scientific papers dating back to 1968
Classic ray tracing shoots virtual view rays into the scene from the camera and traces their paths as they bounce around
With ray tracing, one can achieve a wide variety of complex lighting effects, such as accurate shadows and reflections/refractions from curved surfaces
Achieving these effects with the same precision is difficult if not impossible with a more traditional rendering pipeline
Ray tracing offers a big advance in visual quality, but comes with an expensive price of notoriously slow rendering times






Tracing a single ray requires determining if that ray intersects any one of potentially millions of primitives
This is the basic problem of ray intersection
Many algorithms exist to make this not only feasible, but remarkably efficient
Tracing one ray is a complex problem and requires serious work to make it run at an acceptable speed
Of course, the big problem is the fact that one needs to trace lots of rays to generate a high quality image



Recall that a ray is a geometric entity with an origin and a direction
A ray in a 3D scene would probably use a 3D vector for the origin and a normalized 3D vector for the direction class Ray {
Vector3 Origin;
Vector3 Direction;
};






We start by ‘shooting’ rays from the camera out into the scene
We can render the pixels in any order we choose (even in random order!), but we will keep it simple and go from top to bottom, and left to right
We loop over all of the pixels and generate an initial primary ray (also called a camera ray or eye ray )
The ray origin is simply the camera’s position in world space
The direction is computed by first finding the 4 corners of a virtual image in world space, then interpolating to the correct spot, and finally computing a normalized direction from the camera position to the virtual pixel
Primary ray
Virtual image Camera position






The initial camera ray is then tested for intersection with the 3D scene, which contains a bunch of triangles and/or other primitives
If the ray doesn’t hit anything, then we can color the pixel to some specified ‘background’ color
Otherwise, we want to know the first thing that the ray hits (it is possible that the ray will hit several surfaces, but we only care about the closest one to the camera)
For the intersection, we need to know the position, normal, color, texture coordinate, material, and any other relevant information we can get about that exact location
If we hit somewhere in the center of a triangle, for example, then this information would get computed by interpolating the vertex data

 We will assume that the results of a ray intersection test are put into some data structure which conveniently packages it together class Intersection {
Vector3 Position;
Vector3 Normal;
Vector2 TexCoord;
Material *Mtl; float Distance;
};
// Distance from ray origin to intersection





Once we have the key intersection information (position, normal, color, texture coords, etc.) we can apply any lighting model we want
This can include procedural shaders, lighting computations, texture lookups, texture combining, bump mapping, and more
Many of the most interesting forms of lighting involve spawning off additional rays and tracing them recursively
The result of the lighting equation is a color, which is used to color the pixel







Shadows are an important lighting effect that can easily be computed with ray tracing
If we wish to compute the illumination with shadows for a point, we shoot an additional ray from the point to every light source
A light is only allowed to contribute to the final color if the ray doesn’t hit anything in between the point and the light source
The lighting equation we looked at earlier in the quarter can easily be adapted to handle this, as c lgti will be 0 if the light is blocked c
 m amb
* c amb
  c lgt i
*
 m dif
  i
 m spec
 n
 h i
 s

Obviously, we don’t need to shoot a shadow ray to a light source if the dot product of the normal with the light direction is negative
Also, we can put a limit of the range of a point light, so they don’t have an infinite influence (bending the laws of physics)






Shadow rays behave slightly differently from primary
(and secondary) rays
Normal rays (primary & secondary) need to know the first surface hit and then compute the color reflected off of the surface
Shadow rays, however, simply need to know if something is hit or not
In other words, we don’t need to compute any additional shading for the ray and we don’t need to find the closest surface hit
This makes them a little faster than normal rays




We say that the shadow rays are spawned off of the surface, or we might say that the primary ray spawned off additional shadow rays
When we spawn new rays from a surface, it is usually a good idea to apply a slight adjustment to the origin of the ray to push it out slightly (0.00001) along the normal of the surface
This fixes problems due to mathematical roundoff that might cause the ray to spawn from a point slightly below the surface, thus causing the spawned ray to appear to hit the same surface





Another powerful feature often associated with ray tracing is accurate reflections off of complex surfaces
If we wanted to render a surface as a perfect mirror, instead of computing the lighting through the normal equation, we just create a new reflection ray and trace it into the scene
Remember that primary rays are the initial rays shot from the camera. Any reflected rays (and others, like refracted rays, etc.), are called secondary rays
Reflected rays, like shadow rays should be moved slightly along the surface normal to prevent the ray from re-intersecting the same surface

d n r r
 d

2
  n






If the reflection ray hits a normal material, we just compute the illumination and use that for the final color
If the reflection ray hits another mirror, we just recursively generate a new reflection ray and trace that
In this way, we can render complex mirrored surfaces that include reflections, reflections of reflections, reflections of reflections of reflections…
To prevent the system from getting caught in an infinite loop, it is common to put an upper limit on the ‘depth’ of the recursion. 10 or lower works for most scenes, except possibly for ones with lots of mirrored surfaces
In any case, most pixels will only require a few bounces, as they are likely to hit a non-mirrored surface sooner or later







Surfaces in the real world don’t act as perfect mirrors
Real mirrors will absorb a small amount of light and only reflect maybe 95%-98% of the light
Some reflecting surfaces are tinted and will reflect different wavelengths with different strengths
This can be handled by multiplying the reflected color by the mirror color at each bounce
We can also simulate partially reflective materials like polished plastic, which have a diffuse component as well as a shiny specular component
For a material like this, we would apply the normal lighting equation, including shooting shadow rays, to compute the diffuse component, then add a contribution from a reflection ray to get the final color (the diffuse and specular components should be weighted so as not to violate conservation of energy…)





Ray tracing can also be used to accurately render the light bending in transparent surfaces due to refraction
Often, this is called transmission instead of refraction .
Transmission is a more general term that also includes translucency, but I think the real reason this word is preferred is because ‘reflection’ and ‘refraction’ look too similar
When a ray hits a transparent surface (like glass, or water…), we generate a new refracted ray and trace that, in a similar way as we did for reflection
We will assume that the transmitted ray will obey Snell’s law ( n
1 sin θ
1
= n
2 sin θ
2
), where n refraction for the two materials
1 and n
2 are the index of

d n
1 n
2
θ
1 n z r t r
 d

2
  n z
 n
1 n
2
 d

  n
 t
 z

1
 z
2  n





When light traveling in a material with a high index of refraction hits a material with a low index of refraction at a steep angle, we get a total internal reflection
When this happens, no refraction ray is generated
This effect can be visible when one is scuba diving and looks up at the water surface. One can only see rays refracting to the outside world in a circular area on the water surface above
Total internal reflection can be detected when the magnitude of the z vector is greater than 1, causing the square root operation to become undefined n d n
1 n
2
θ
1 z r z
 n
1 n
2
 d

  n
 t
 z

1
 z
2  n







When light hits a transparent surface, we not only see refraction, but we get a reflection off of the surface as well
Therefore, we will actually generate two new rays and trace both of them into the scene and combine the results
The results of an individual traced ray is a color, which is the color of the light that the ray ‘sees’
This color is used as the pixel color for primary rays, but for secondary rays, the color is combined somehow into the final pixel color
In a refraction situation, for example, we spawn off two new rays and combine them according to the Fresnel equations, provided in the last lecture
The Fresnel equations describe how the transmitted (refracted) ray will dominate when the incoming ray is normal to the surface, but the reflection will dominate when the incoming ray is edge-on

Camera
Normal
Primary ray
Reflection ray
Transmission ray



The Fresnel equations can be used to determine the proportion of the light reflected
( f r
) and transmitted ( f t
) when a ray hits an interface between two dielectrics (like air and water)
They describe separate formulas for the parallel and perpendicularly polarized light, but these are usually averaged into a single set of values r par
 n
2
( n n
2
( n
 d
 d )
)

 n
1
( n
 t n
1
( n
 t )
) r perp
 n
1
( n
 d n
1
( n
 d )
)

 n
2
( n
 t n
2
( n
 t )
) f f t r

1
2

1 .
0
(
2 r par
 f r
 2 r perp
)

 The classic ray tracing algorithm includes features like shadows, reflection, refraction, and custom materials
 A single primary ray may end up spawning many secondary and shadow rays, depending on the number of lights and the arrangement and type of materials
 These rays can be thought of as forming a tree like structure






One of the key components of a ray tracer is the system that determines what surface the ray hits
A typical 3D scene may have well over 1,000,000 primitives
As usual, triangles tend to be the primitive of choice, but one advantage of a ray tracer is that one can intersect rays with more complex surfaces such as spheres, Bezier patches, displacement mapped surfaces, fractals, and more
Sometimes, complex primitives are simply tessellated into triangles in a pre-rendering phase, and then just ray traced as triangles
Alternately, it is possible to ray trace complex surfaces directly, or to use demandbased schemes that don’t tessellate an object until a ray comes nearby





We will say that our scene is made up of several individual ‘objects’
For our purposes, we will allow the concept of an object to include primitives such as triangles and spheres, or even collections of primitives or other objects
In order to be render-able, an object must provide some sort of ray intersection routine
We will define a C++ base class object as: class Object { public: virtual bool IntersectRay(Ray &r,Intersection &isect);
};



The idea is that we can derive specific objects, like triangles, spheres, etc., and then write custom ray intersection routines for them
The ray intersect routine takes a ray as input, and returns true if the object is hit and false if it is missed
If the object is hit, the intersection data is filled in into the isect class


Let’s see how to test if a ray intersects a sphere
 The ray has an origin at point p and a unit length direction u , and the sphere has a center c and a radius r p u c r





The ray itself is the set of points p + αu , where α ≥0
We start by finding the point q which is the point on the ray-line closest to the center of the sphere
The line qc must be perpendicular to vector u, in other words, ( q c )· u =0, or ( p + αu -c) · u =0
We can solve the value of α that satisfies that relationship: α =-( p c )· u , so q = p -(( p c )· u ) u p u q c





Once we have q, we test if it is inside the actual sphere or not, by checking if | q c | ≤ r
If q is outside the sphere, then the ray must not miss
If q is inside the sphere, then we find the actual point on the sphere surface that the ray intersects
We say that the ray will hit the sphere at two points q
1 and q
2
: q
1
= p +( α a ) u ) q
2
= p +( α + a ) u ) where a =sqrt(r 2 -| q c | 2 )


If
α
a
≥0, then the ray hits the sphere at q
1
, but if it is less than 0, then the actual intersection point lies behind the origin of the ray
In that case, we check if α+a ≥0 to test if q
2 is a legitimate intersection q
1 q
2 p u q c





There are several ways to formulate the raysphere intersection test
This particular method is the one provided in the book
As a rule, one tries to postpone expensive operations, such as division and square roots until late in the algorithm when it is likely that there will be an intersection
Ideally, quick tests can be performed at the beginning that reject a lot of cases where the ray is far away from the object being tested







A plane is defined by a normal vector n and a distance d , which is the distance of the plane to the origin
We test our ray with the plane by finding the point q which is where the ray line intersects the plane
For q to lie on the plane it must satisfy d = q · n = p · n + αu · n
We solve for α:
α =( d p · n )/( u · n )
However, we must first check that the denominator is not 0, which would indicate that the ray is parallel to the plane
If α ≥0 then the ray intersects the plane, otherwise, the plane lies behind the ray, in the wrong direction




To intersect a ray with a triangle, we must first check if the ray intersects the plane of the triangle
If we are treating our triangle as one-sided, then we can also verify that the origin of the ray is on the outside of the triangle
Once we know that the ray hits the plane at point q , we must verify that q lies inside the 3 edges of the triangle

Does segment ab intersect triangle v p u v
2
• q v
0 v
1




Reduce to 2D: remove smallest dimension
Compute barycentric coordinates v
2 q' = q v
0 e
1
= v
1
v
0 e
2
= v
2
v
0
=( q ' · e
2)
/( e
1
· e
2
) v
0
β =( q ' · e
1)
/( e
1
· e
2
)
Reject if α<0, β <0 or α+ β >1
β
α q v
1







Complex scenes can contain millions of primitives, and ray tracers need to trace millions of rays
This means zillions of potential ray-object intersections
If every ray simply looped through every object and tested if it intersected, we would spend forever just doing loops, not even counting all of the time doing the intersection testing
Therefore, it is absolutely essential to employ some sort of acceleration structure to speed up the ray intersection testing
An acceleration structure is some sort of data structure that groups objects together into some arrangement that enables the ray intersection to be sped up by limiting which objects are tested
There are a variety of different acceleration structures in use, but most of the successful ones tend to be based on some variation of hierarchical subdivision of the space around the group of objects









The basic concept of a bounding volume hierarchy is a complex object in a hierarchy of simpler ones
This works much like the hierarchical culling we looked at in the scene graph lecture
For example, if one were using spheres as their bounding volume, we could enclose the entire scene in one big sphere
Within that sphere are several other spheres, each containing more spheres, until we finally get to the bottom level where spheres contain actual geometry like triangles
To test a ray against the scene, we traverse the hierarchy from the top level
When a sphere is hit, we test the spheres it contains, and ultimately the triangles/primitives within
In general, a bounding volume hierarchy can reduce the ray intersection time from
O( n ) to O(log n ), where n is the number of primitives in the scene
This reduction from linear to logarithmic performance makes a huge difference and makes it possible to construct scenes with millions of primitives




The sphere hierarchy makes for a good example of the concept, but in practice, sphere hierarchies are not often used for ray tracing
One reason is that it is not clear how to automatically group an arbitrary set of triangles into some number of spheres, so various heuristic options exist
Also, as the spheres are likely to overlap a lot, they end up triggering a lot of redundant intersection tests





The octree starts by placing a cube around the entire scene
If the cube contains more than some specified number of primitives (say, 10), then it is split equally into 8 cubes, which are then recursively tested and possibly resplit
The octree is a more regular structure than the sphere tree and provides a clear rule for subdivision and no overlap between cells
This makes it a better choice usually, but still not ideal








The KD tree starts by placing a box (not necessarily a cube) around the entire scene
If the box contains too many primitives, it is split, as with the octree
However, the KD tree only splits the box into two boxes, that need not be equal
The split can take place on the x, y, or z place at some arbitrary point within the box
This makes the KD tree a little bit more adaptable to irregular geometry and able to customize a tighter fit
In general, KD trees tend to be pretty good for ray tracing
Their main drawback is that the tree depth can get rather deep, causing the ray intersection to spend a lot of time traversing the tree itself, rather than testing intersections with primitives





The BSP tree ( binary space partitioning ) is much like the
KD tree in that it continually splits space into two (not necessarily equal) halves
Unlike the KD tree which is limited to xyz axis splitting, the BSP tree allows the splitting plane to be placed anywhere in the volume and aligned in any direction
This makes it a much more difficult problem to choose the location of the splitting plane, and so many heuristics exist
In practice, BSP trees tend to perform well for ray tracing, much like KD trees





One can also subdivide space into a uniform grid, instead of hierarchically
This is fast for certain situations, but gets too expensive in terms of memory for large complex scenes
It also tends to loose its performance advantages in situations where primitives have a large variance in size and location (which is common)
As a result, they are not really a practical general purpose acceleration structure for ray tracing






One can also make a hierarchical grid
Start with a uniform grid, but subdivide any cell that contains too many primitives into a smaller grid
An octree is an example of a hierarchical grid limited to
2x2x2 subdivision
A more general hierarchical grid could support subdivision into any number of cells
Hierarchical grids tend to perform very well in ray tracing, especially for highly detailed geometry of relatively uniform size (such as the triangles in a tessellated surface)






All of the acceleration structures we looked at store some geometry and provide a function for intersecting a ray
In other words, they are really just a more complex type of primitive themselves
We can derive acceleration structures off of our base Object class, just like we did for Spheres and Triangles
Also, acceleration structures can be designed so that they store a bunch of generic Objects themselves, and so one could build an acceleration structure that contains a bunch of triangles, and then place that acceleration structure within a larger acceleration structure, etc.
This provides a nice, consistent way to represent scenes, similar to the scene graph concept we covered in the lecture on realtime scene management class KDTree:public Object { public: bool IntersectRay(Ray &r,Intersection &isect);
};





In 1984, an important modification to the basic ray tracing algorithm was proposed, known as distributed ray tracing
The concept basically involved shooting several distributed rays to achieve what had previously been done with a single ray
The goal is not to simply make the rendering slower, but to achieve a variety of ‘soft’ lighting effects such as antialiasing, camera focus, soft-edge shadows, blurry reflections, color separation, motion blur, and more
As the term ‘distributed’ tends to refer to parallel processing in modern days, the distributed ray tracing technique is now called distribution ray tracing , and the term ‘distributed’ is reserved for parallel ray tracing, which is also an important subject





One nice visual effect we can achieve with distribution ray tracing is soft shadows
Instead of treating a light source as a point and shooting a single ray to test for shadows, we can treat the light source as a sphere and shoot several rays to test for partial blocking of the light source
If 15% of the shadow rays are blocked, then we get 85% of the incident light from the light source
In lighting terminology, the completely shadowed region is called the umbra and the partially shadowed region is called the penumbra






The soft shadow technique enables us to define lights in a much more complex way than we have previously
We can now use any geometry to define a light, including triangles, patches, spheres, etc.
To determine the incident light, we shoot several rays towards the light source, distributed across the surface and weighted according to the surface area of the sample and the direction of the average normal
Larger light sources create softer, diffuse shadows, while smaller light sources cause sharp, harsh shadows
Larger light sources also require more rays to adequately sample the shadows, making area lights a lot more expensive than point lights. Inadequate sampling of the light source can cause noise patterns to appear in the penumbra region, known as shadow aliasing





We can render blurry or glossy reflections by creating several reflection rays instead of just one
The rays can be distributed around the ideal reflection direction
Blurry surfaces will causes a wider distribution (and require more rays), while more polished surfaces will have a narrow distribution
The same concept can apply to refraction in order to achieve rendering of unpolished glass




For good quality antialiasing, we will can shoot several primary rays for each pixel
We covered several different antialiasing patterns in an earlier lecture, but recall that the supersampled, area-weighted, jittered
Gaussian distribution was chosen as one of the best overall sampling schemes
As we have complete control over the direction of every ray, we can use any sampling pattern we choose, as well as any number of rays we choose








The blurring caused by a camera lens being out of focus is due to the lens’ limited depth of field
In computer graphics, the term ‘depth of field’ usually refers to the general process of rendering images that include a camera blurring effect
A lens will typically be set to focus on objects at some distance away, known as the focal distance
Objects closer or farther than the focal distance will be blurry, and the blurriness increases with the distance to the focal plane
Depth of field can be rendered with distribution ray tracing by distributing the primary rays shot from the camera
Rays area distributed across a virtual aperture , which represents the
(usually circular) opening of the lens
The larger the aperture, the more pronounced the blurring effect will be. A pinhole camera has an aperture size of 0, and therefore, will not have any blurring due to depth of field





As if soft shadows, antialiasing, depth of field, and blurry reflections and refractions weren’t enough… we can also use distribution ray tracing to include the effect from motion blur
We do this by distributing rays in time
Therefore, we need to know the starting and ending transformations for all of the objects in the scene (and the lights & camera)
In fact, we can even consider objects moving along more complex paths, so we can motion blur bounces as well




Ray tracing had a big impact on computer graphics in
1980 with the first images of accurate reflections and refractions from curved surfaces
Distribution ray tracing had an even bigger impact in
1984, as it re-affirmed the power of the basic ray tracing technique and added a whole bunch of sophisticated effects, all within a consistent framework
Previously, techniques such as depth of field, motion blur, soft shadows, etc., had only been achieved individually and by using a variety of complex, hacky algorithms









If ray tracing is slow, then distribution ray tracing must be considerably slower
Now, instead of one or two splits per level in our recursion, we are have to shoot dozens or even hundreds of rays to achieve some of these effects
This can cause an exponential expansion in the number of rays
The good news is that we can combine these features so that we still only need to shoot a small number of primary rays per pixel
For example, we can shoot 16 rays in a 4x4 antialiasing pattern, where each ray has a random distribution in time and in the camera aperture
Each of these rays only needs to spawn a few reflection or shadow rays, as the results will be blended with 15 other samples
Still, we end up with lots and lots of rays and potential for exponential problems in scenes with a lot of soft or blurry features
This problem is at least partially addressed with path tracing which is one of the techniques for global illumination that we will see in the next lecture