COMPUTER GRAPHICS COURSE

Ray Tracing

Georgios Papaioannou - 2015

RAY TRACING PRINCIPLES
 What is ray tracing?

• A general mechanism for sampling paths of light in a 3D scene

• We will use this mechanism in path tracing

• Ray Casting mechanism:

– Rays are cast from a point in space towards a specified direction
– Rays are intersected with geometry primitives
– The closest intersection is regarded as a ray “hit”
– Lighting or other attributes are evaluated at the hit location
 Simple Ray Casting - Appel’s Method

• For each image pixel, a ray (line segment) is cast from the
viewpoint, crosses the pixel and is directed toward the scene
• For hit points:
– Send (at least) one ray to each
light source and check visibility
(shadow)
– Shade the point using a local
illumination model
 Whitted-style Ray Tracing (1)

• For each image pixel, a ray (line segment) is cast from the
viewpoint, crosses the pixel and is directed toward the scene
• The ray hits the objects, is absorbed or deflected in order to
gather what the observer would “see” through each pixel
• This is a recursive algorithm that spawns new rays at each hit
point
Whitted-style Ray Tracing (2)
 Why not Trace Rays from the Lights? (1)

• Infinite rays leave a light source but only a small number lands
on the viewport
– Even fewer when a pinhole camera is considered
– Extremely low probability to hit  Computationally intractable
 Why not Trace Rays from the Lights? (2)

• In practice, many path tracing variants, which use the ray

tracing mechanism to form the light paths, do trace rays from
both the camera and the light source domain

Source: http://indigorenderer.com
 More than Direct Illumination

• Ray tracing is an elegant recursive algorithm

• The first (primary) rays spawned hit the nearest surface: same
result as in direct rendering
• Secondary rays can be spawned from the intersection point to
track indirect illumination
• Secondary rays capture:
– Shadows (inherent part of ray-tracing, no special algorithm
required)
– Reflections (light bouncing off surfaces)
– Refracted (transmitted) light through the objects
Tracing Rays: Level 0 (Primary Rays)

• Determine ray from viewpoint and

pixel center
• Intersect ray with scene
• Calculate local illumination model
Tracing Rays: Level 1

• Determine shadow ray from light

and intersection
• Test light source visibility
• Calculate and cast reflection ray
• Calculate and cast refraction ray
Tracing Rays: Level 2

• Determine shadow ray from light

and intersection
• Test light source visibility
• Calculate and cast refraction ray
(exiting)
• Estimate local illumination
Tracing Rays: Level 3

• Determine shadow ray from light

and intersection
• Test light source visibility
• Calculate and cast reflection ray
• Estimate local illumination
Tracing Rays: Level 4

• Determine shadow ray from light

and intersection
• Test light source visibility (in
shadow here) – ray intersects
opaque geometry
 Who Determines What Rays to Spawn?

• Material properties:
– Reflectance
– Incandescence
– Gloss
– Permeability a
– Index of refraction n
– …
• Number, size and type of lights
Resulting Color

Spawn Gather
 The Basic Ray Tracing Algorithm
Color raytrace( Ray r, int depth, Scene world, vector <Light*> lights )
{ Ray *refl, *tran;
Color color_r, color_t, color_l;

// Terminate if maximum recursion depth has been reached.

if ( depth > MAX_DEPTH ) return backgroundColor;
// Intersect ray with scene and keep nearest intersection point
int hits = findClosestIntersection(r, world);
if ( hits == 0 ) return backgroundColor;
// Apply local illumination model, including shadows
color_l = calculateLocalColor(r, lights, world);
// Trace reflected and refracted rays according to material properties
if (r->isect->surface->material->k_refl > 0)
{ refl = calculateReflection(r);
color_r = raytrace(refl, depth+1, world, lights);
}
if (r->isect->surface->material->k_refr > 0)
{ tran = calculateRefraction(r);
color_t = raytrace(tran, depth+1, world, lights);
}
return color_l + color_r + color_t;
}
 Ray Tracing Results (1)

Source: http://hof.povray.org/images/ChristmasBaubles.jpg
 Ray Tracing Results (2)

Simple ray-traced scene rendered at 60 frames per second on a modern GPU at 1080p
Source: NVIDIA OptiX SDK
 Comments

• Conventional primary and secondary ray – scene intersections

must return the closest hit to the ray origin
• If reflection / refraction coefficients are (near) zero, no
secondary rays are spawned
• If no surface is hit, the background color is returned
• Shadow determination is embedded in the local illumination
calculation stage
• A maximum recursion depth is forced upon the algorithm;
Necessary for complex scenes
 Termination Criteria

• Maximum recursion depth is reached

• Zero reflectivity / transmission coefficients
• Ray contribution too small to be of significance:
Attenuation due to participating medium density
Effect of Ray Tracing Depth on Images
 Ray Data Structures – Requirements

• Degradation effects (absorption, scattering, splitting):

– “strength” indicator (opposite of attenuation)
– Optionally, recursion depth
• Distance sorting of hit points:
– Avoid keeping all intersections and post-sort results
– Keep nearest intersection point or
– Cache distance to nearest hit point
• Local properties of hit point:
– Need to keep track of hit material, primitive and local
attributes (e.g. normal)
 A Ray as a Data Structure - Minimum
class ray
{
public:
ray(const vec3 & start, const vec3 & direction);
void transform(const mat4 &m);

vec3 origin; Here, position is indirectly

vec3 dir; calculated from t
vec3 n_isect;
real t; // real: defined as float or double
void * payload;
}; Pointer to an existing
structure (e.g. a primitive)
that holds the local
attributes associated to
the hit point
 A Ray as a Data Structure - Extended
class ray
{
public:
ray(void);
ray(const vec3 & start, const vec3 & direction);
void transform(const mat4 &m);

vec3 origin;
vec3 dir;
int depth;
vec3 p_isect;
vec3 n_isect;
vec3 barycentric;
real t;
real strength;
bool hit;
bool inside;
class primitive *hit_primitive;
void *payload;
};
 Ray – Scene Intersection: Primitives

• A primitive in ray tracing is any mathematical entity that can

define a line-primitive equation (intersection points)
– Polygons
– Thresholded density (volume) data
– Parametric surfaces
– Analytical surfaces (e.g. spheres)
– General equations (e.g. fractals)
– Set operations (Constructive Solid
Geometry)

Sources: http://paulbourke.net/geometry/blob/ http://www.subblue.com/blog/2009/12/13/mandelbulb

 Ray Tracing Results (3)

Ray tracing using only geometric solids (with CSG in POVRAY )

 Ray – Scene Intersection

• A naïve test exhaustively tests a ray against every

primitive in the scene
• Ray – primitive intersections are the most frequent
operations in ray tracing
• We try to minimize their complexity and number:
– Hierarchical data structure acceleration
– Early visibility culling
– Frequent tests are performed with low-complexity
primitives  search refinement
– Parallel execution: Ray tracing is inherently highly parallel
at many levels
 Nearest Hit Determination
int findClosestIntersection(Ray r, Scene world)
{
int hits=0;
Ray r_temp = r;
r.t = FLT_MAX;
for ( j=0; j<world.numObjects(); j++ )
for ( k=0; k<world.getObject(j)->numPrims(); k++ )
{
Primitive *prim = world.getObject(j)->getPrim(k);
prim->intersect(r_temp);
hits++;

if ( r_temp.t < r.t )

r = r_temp;
} All intersectable entities
return hits; here are derived from the
} Primitive class and
override the intersect()
method
 Shadows

• An intersection point is in shadow when the direct path to a

light source is obstructed by a surface

in shadow

lit
 Shadow Rays

• Cast “shadow rays” toward each light source during the local
color estimation
• Shadow rays are cheaper:
– Once light is blocked by a primitive, the search stops
– No sorting is required
• Occluding geometry does not necessarily completely block or
allow light through
– Modulate the ray strength according to occluder
transparency
– Stop if strength becomes too low (in shadow)
 Shadow Determination (1)

Color calculateLocalColor( Ray r, vector<Light*> lights,

Scene world ) // point lights are assumed here
{
int i, j, k;
Color col = Color(0); // black

// For all available lights, trace array towards them

for ( i=0; i<lights.size(); i++ )
{
vec3 dir = normalize(lights[i]->pos-r.p_isect);
Ray shadowRay = Ray(r.p_isect, dir);

float strength = 1.0f;

// Filter the light as it passes through the scene
<SEE NEXT SLIDE>

if (strength>0)
col += strength * localShadingModel(r,prim,lights[i]->pos);
}
return col;
}
 Shadow Determination (2)

...
// Filter the light as it passes through the scene
for ( j=0; j<world.numObjects(); j++ )
for ( k=0; k<world.getObject(j)->numPrims(); k++ )
{
Primitive *prim = world.getObject(j)->getPrim(k);
if (prim->intersect(r));
strength *= prim->material->getTransparency(r);

// Termination criterion: light almost cut off

if ( strength < 0.002 )
{
strength=0;
break;
}
}
...
RAY GENERATION
 Shooting Rays – Primary Rays (1)

wv  d tan  hv  wv / a
r r r r  h r r  
pUL  c  d  n  wv u  hv v  UL
p  c  d n 
 w  v  u  tan  
   
r 2wv r r 2hv r
u  u v   v
w h
 Shooting Rays – Primary Rays (2)

• The center of each (i, j) pixel in WCS is:

 1 r  1 r
p  pUL   i    u   j    v
 2  2
• And the corresponding ray (direction) that passes through it is
given by:
r p -c
r
p -c
 Primary Rays - Ray Segment Definition

• Starting point:
 1 r  1 r
– Either p  pUL   i    u   j   v (planar near surface)
 2  2
r
– Or pstart  c  n  r (Spherical near surface – Can be zero!)

r
• Arbitrary ray point: q  q(t )  pstart  t r

• t is the (signed) distance from the origin as ray vector is normalized

 Primary Rays - Clipping Distances

• Depth buffer and perspective projection require a near and a

far clipping distance (plane)
• In ray tracing, depth sorting is handled by ray hit sorting so no
special ranges are required
• Distance from viewpoint can take arbitrary values (even
negative – back of viewer)
• Far clipping distance determined by numerical limits
• Near clipping distance can be zero
• Depth resolution same as floating point precision
 Shooting Rays - Secondary Rays

• Origin = Last intersection point

• Direction =
– Reflected vector
– Refracted vector
r li  q
– Vector to i-th light source: r
li  q
 Secondary Rays – Coincidence Pitfall

• Sec. rays can and will intersect with originating surface point
(self intersection)
• Fix:
– Offset the origin along its direction before casting
 Reflection Direction

𝐫Ԧ𝑟 = 2𝐭Ԧ −Ԧ𝐫𝑖

r r
𝐭Ԧ = 𝐫Ԧ𝑖 − 𝑝𝑟𝑜𝑗𝐧 𝐫Ԧ𝑖 = t t

𝐫Ԧ𝑖 + 𝐧cos𝜃𝑖 =

𝐫Ԧ𝑖 − 𝐧(𝐧Ԧ𝐫𝑖 ) ⇒

𝐫Ԧ𝑟 = 𝐫Ԧ𝑖 − 2𝐧(𝐧Ԧ𝐫𝑖 )

 Refraction – Index of Refraction

• When light enters a dielectric medium,

its phase velocity changes (const.
frequency)
• The ratio of its phase velocity in the
medium and c (vacuum) is the IOR n:
n  c
• n≈1 for thin air
• n>1 for transparent materials
 Refraction – Snell’s Law

• At the interface between 2 media

with IOR n1 and n2, the ray is bent
according to the law:

sin  t n1

sin  i n2
Refraction Direction (1)
 Refraction Direction (2)

• From Pythagorean theorem:

cos t  1  sin 2 t 
n12
1  2 sin 2 i 
n2
n12
1  2 1  cos 2 i 
n2

• Using dot product instead of cosine:

 Refraction Direction (3)

• Critical angle:
• When 𝑛1 > 𝑛2 , reflection may occur instead of refraction for
incident angles higher than a certain threshold 𝜃𝑐
• This is called total internal reflection and can be easily
observed underwater, when looking upwards almost parallel
to the surface

Becomes negative for incident

directions beyond the critical angle
 Ray Transformations

• When rays are intersected with moving geometry, BVH trees,

or other elements with parameters defined in a local
coordinate system:
– It is more efficient to transform the ray instead of the
object! (why?)
– Example: OBB/BV hierarchies (common structure for scene
graphs)
1 1 r
• q = M  q  M  Object.RayIntersection (M  p, M  r )
– Ray expressed in the local reference frame
– The result is expressed back in WCS
 Ray Transformations - Example

Transformed sphere
according to matrix 𝐌. Local CS
Sphere is distorted
𝐌 −1 ∙ray
q'
y

q
Perform easy-to-compute
ray ray-sphere intersection

WCS x

q = 𝐌 ∙ q’
RAY TRACING ACCELERATION TECHNIQUES
 Basic Acceleration Concepts

Source: https://graphics.stanford.edu/courses/cs348b-05/lectures/lecture3/raytrace_ii.pdf
Bounding Volumes
a. Axes-aligned bounding
box (AABB)
b. Oriented bounding box
(OBB)
c. BV hierarchy (BVH)
d. Bounding slabs
 Bounding Volumes – Pros & Cons

• AABB:
– Easy to implement and initialize
– Fast test, no ray transformations required
– Can leave too much void space  degraded pruning performance
• OBB:
– Can be costly to initialize (e.g. PCA algorithm)
– Fast test, ray transformation required
– Ideal for animated hierarchies (no recalculation of extents required)
– Tighter fitting than AABB
• Bounding Slabs:
– Very efficient, even less void space
– More computationally expensive than AABB/OBB
Ray - Scene Graph/BVH Intersection
 Spatial Subdivision Acceleration (1)

• Primitives can be organized into “bins”, according to rough

position in space
• When a ray is cast, it registers the bins it passes through and
only tests primitives inside those bins

• Spatial subdivision structures can be local to aggregate scene

nodes (groups)
• And nested
– Use ray transformations to go from one local coordinate system to the
next
Spatial Subdivision Acceleration (2)

Uniform grid Visit cells using a 3D DDA method

 Hierarchical Spatial Subdivision

The spatial subdivision bins (cells) can be hierarchically

organized too.

Source: https://graphics.stanford.edu/courses/cs348b-05/lectures/lecture3/raytrace_ii.pdf
 Hierarchical Spatial Subdivision

Recursive in-order traversal: rays are tested with

subspaces of a splitting plane (binary subdivision)

Source: https://graphics.stanford.edu/courses/cs348b-05/lectures/lecture3/raytrace_ii.pdf
 Octree

• Common structure is the octree:

• Subdivide space in 8 cells:
– Up to max depth
– Until cell contains no
primitives
 K-d Trees

• Typically K=3 in graphics (3D)

• With the K-d tree, 2 things must be determined at
each level:
– Which axis to split  usually the longest
– Where to set the split
• Median cut
• Midpoint
• SAH (surface area heuristic)
 K-d Tree Construction Example
1 2 3 4
C C C C
B B B B
E E E E
D D D D
A A A A

A, B, C, A, B,
E E E
D, E C, D

A, B, B, B,
D C C
A, A,
B D
 Complexity Analysis of a Split (1)

• To decide to split a cell, the cost of not splitting it 𝐶𝑁𝑆

should be greater than the cost of using a split 𝐶𝑆
• For 𝑁𝑂 primitives in the cell, each with intersection
cost 𝐶𝑂 , the cost of using the cell undivided is: 𝐶𝑁𝑆 =
𝑁𝑂 ∙ 𝐶𝑂

• The probability that a ray hits a convex shape A

completely within another convex shape B is:
𝑆𝐴(𝐴)
𝑃𝐴 = , where 𝑆(𝑋) the surface area
𝑆𝐴(𝐵)
 Complexity Analysis of a Split (2)

• Consider only one splitting axis and a parameter 𝑏 ∈ 0,1 ,

determining where the split occurs
– For 𝑏=1/2: the spatial median, i.e. in the middle
• The maximum traversal cost (no intersections found, no early
termination) of the split cell is the weighted sum of the cost
for the two new cells:
• 𝐶𝑆 𝑏 = 𝑃𝐿 𝑏 𝑁𝐿 𝑏 𝐶𝑂 + 𝑃𝑅 𝑏 𝑁𝑅 𝑏 𝐶𝑂 =

𝑆𝐴(𝐿) 𝑆𝐴(𝑅)
𝑁𝐿 𝑏 𝐶𝑂 + 𝑁𝑅 𝑏 𝐶𝑂
𝑆𝐴(𝐿 ∪ 𝑅) 𝑆𝐴(𝐿 ∪ 𝑅)
 Complexity Analysis of a Split (3)

• 𝑆𝐴(𝐿 ∪ 𝑅) is the surface of the un-split cell

• Where 𝑁𝐿 𝑏 , 𝑁𝑅 𝑏 are the number of primitives in the left
and right part of the subdivided cell
• Note that 𝑁𝐿 𝑏 + 𝑁𝑅 𝑏 ≠ 𝑁𝑂 in general, as primitives may
cross the split boundary
 Surface Area Heuristic (1)

• Determines a splitting plane (and potentially axis,

too), by minimizing the above cost function 𝐶𝑆
• Facts:
– Discontinuous function
– Optimal cut between spatial median and midpoint
• Two options:
– Sort primitive bounds per axis, locate median and test
bounds between median and midpoint
– Greedily test all bounds
• Number of bounds: 2𝑁𝑂 or 6𝑁𝑂 for concurrent axis
selection
 Surface Area Heuristic (2)

Midpoint Splits

http://www.keithlantz.net/2013/04/kd-tree-construction-using-the-surface-area-heuristic-stack-based-traversal-and-the-hyperplane-separation-theorem/
 Surface Area Heuristic (3)

Median Splits

http://www.keithlantz.net/2013/04/kd-tree-construction-using-the-surface-area-heuristic-stack-based-traversal-and-the-hyperplane-separation-theorem/
 Surface Area Heuristic (4)

SAH Splits

http://www.keithlantz.net/2013/04/kd-tree-construction-using-the-surface-area-heuristic-stack-based-traversal-and-the-hyperplane-separation-theorem/
INTERSECTION TESTS
 Intersection Tests: Ray - Plane

• If the plane equation is:

• We substitute point p by the line definition:
r
p1 n
• So: p2

• If instead of we use a normalized vector, t is the signed

distance along the ray
 Intersection Tests: Ray - Triangle (1)

• Barycentric triangle coordinates:

– Any point in the triangle can be expressed as a weighted
sum of the triangle vertices (affine combination):

q(u, v, w)  wv 0  uv1  vv 2 ,
u v  w 1
 Intersection Tests: Ray - Triangle (2)
• Requiring intersection point in triangle:
r
p  td  (1  u  v) v0  uv1  vv2

• And in the form of a linear system (3 unknowns):

t 
r
 d v1  v 0
 v 2  v 0  u   p  v 0 
 v 

• We solve it for t, u and v

• If u, v and 1- u - v ≤ 1, then hit point inside triangle

• See [RTI] for an efficient implementation of the above

 Intersection Tests: Ray – Sphere (1)

• A Ray intersects a sphere if:

– Line – sphere equation has 1 root and 0 ≤ t (otherwise the
ray points away from the sphere)
– Line – sphere equation has 2 roots:
• 2 negative: ray points away (no intersection)
• 1 positive, 1 negative: positive root defines the intersection point
• 2 positive roots, smallest one corresponds to entry point
 Intersection Tests: Ray – Sphere (2)

• Combining the sphere parametric equation

with the line parametric equation: we get:

where is a vector from the center of the sphere to

the ray origin
 Intersection Tests: Ray – Sphere (3)

• This is a normal quadratic equation for t of the form:

at 2  2bt  c  0
r r r r r r
where: a  d  d, b  m d, c m m  r2
• The discriminant b2  ac specifies the roots and
corresponding intersection points:
– D<0: No intersection
– D=0: One intersection
b  D
– D>0: 2 intersection points: t 
a
 Deficiencies of Simple Ray Tracing

• Marginally interactive method, even with

optimizations only for simple scenes
• Extremely (and unnaturally) crisp and polished
images
– Ideal specular (mirror) reflection and transmission
– Natural surfaces and media are not “ideal”
• No other light transport event is modelled
 Contributors

• Georgios Papaioannou

References
[RTI]: Fast, Minimum Storage Ray/Triangle Intersection , Möller &
Trumbore. Journal of Graphics Tools, 1997