Lesson 7

Stochastic Ray Tracing

Introduction
Supersampling
Soft Shadows
Glossy Surface
Motion Blur
Depth of Field
Translucency

Introduction

primary ray

shadow ray

Why does ray tracing look obviously computer generated ?

Crisp images are too "perfect":

Surfaces are perfectly shiny
Glass is perfectly clear
Everything in perfect focus
Every object is completely still
Even the shadows have perfect silhouettes
But . . . up close, edges are jagged

Introduction

primary ray

shadow ray

transmitted ray

reflected ray

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Average results together

Introduction

primary ray

shadow ray

transmitted ray

reflected ray

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Average results together

Multiple primary rays through a pixel

Supersampling: distribute rays spatially
Motion blur: distribute rays temporally
Depth of field: distribute rays through a lens

Introduction

primary ray

shadow ray

transmitted ray

reflected ray

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Average results together

Multiple primary rays through a pixel

Supersampling: distribute rays spatially
Motion blur: distribute rays temporally
Depth of field: distribute rays through a lens

Multiple shadow rays to sample an area light

Soft Shadows

Introduction

primary ray

shadow ray

transmitted ray

reflected ray

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Average results together

Multiple primary rays through a pixel

Supersampling: distribute rays spatially
Motion blur: distribute rays temporally
Depth of field: distribute rays through a lens

Multiple shadow rays to sample an area light

Soft Shadows

Multiple reflection rays

Glossy surfaces: blurry reflection, rough specular

Introduction

primary ray

shadow ray

transmitted ray

reflected ray

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Average results together

Multiple primary rays through a pixel

Supersampling: distribute rays spatially
Motion blur: distribute rays temporally
Depth of field: distribute rays through a lens

Multiple shadow rays to sample an area light

Soft Shadows

Multiple reflection rays

Glossy surfaces: blurry reflection, rough specular

Multiple refraction rays

Translucency

Supersampling

Instead of point sampling the color of a pixel with a ray, we cast multiple rays from eye (primary rays) through different parts of one pixel and average down the results

Supersampling

Instead of point sampling the color of a pixel with a ray, we cast multiple rays from eye (primary rays) through different parts of one pixel and average down the results

For example, cast $n\times n$ sub-pixels rays, and average the results together:

\[c_{pixel}=\frac{1}{n^2}\sum^{n-1}_{i=0}\sum^{n-1}_{j=0}c_{subpixel(i,j)}\]

Supersampling

Super-sampling does not eliminate aliasing,
it simply pushes it to higher frequencies

Supersampling captures more high frequencies, but frequencies above the supersampling rate are still aliased
Fundamentally, problem is that the signal is not band-limited $\Rightarrow$ aliasing happens

Other than regular, fixed sampling pattern, sampling can also be stochastic
(a.k.a., random, probabilistic, or Monte Carlo)

Non-fixed Sampling Patterns

Supersampling

Samples taken at non-uniformly spaced random offsets

Replaces low-frequency aliasing pattern by noise, which is less objectionable to humans

Stochastic Supersampling

uniform sampling

stochastic
sampling

rt::CSamplerStratified(
  nSamples = 4, 
  isRenewable = true, 
  jitter = false
)

rt::CSamplerRandom(
  nSamples = 4, 
  isRenewable = true
)

Supersampling

Samples taken at non-uniformly spaced random offsets

Replaces low-frequency aliasing pattern by noise, which is less objectionable to humans

However, with random sampling, we could get unlucky, e.g., all samples in one corner

Over 80% of the samples are black while the pixel should be light grey

Stochastic Supersampling

uniform sampling

stochastic
sampling

rt::CSamplerStratified(
  nSamples = 4, 
  isRenewable = true, 
  jitter = false
)

rt::CSamplerRandom(
  nSamples = 4, 
  isRenewable = true
)

Supersampling

To prevent clustering of the random samples, divide $[0, 1)^2$ domain (pixel) into non-overlapping regions (sub-pixels) called strata

Take one random sample per stratum

Jittered sampling is stratified sampling with per-stratum sample taken at an offset from the center of each stratum:

One sample per stratum
Randomly perturb the sample location
Size of perturbation vector limited by the subpixel distance
Patented by Pixar!

Stratified Sampling

uniform sampling

stratified
sampling

rt::CSamplerStratified(
  nSamples = 4, 
  isRenewable = true, 
  jitter = false
)

rt::CSamplerStratified(
  nSamples = 4, 
  isRenewable = true, 
  jitter = true
)

Supersampling

1 ray per pixel

Stratified Sampling

Supersampling

16 rays per pixel (uniform)

Stratified Sampling

Supersampling

16 rays per pixel (jittered)

Stratified Sampling

Supersampling

Motion Blur

Aliasing happens in time as well as in space

The sampling rate is the frame rate: 30Hz (NTSC),
25Hz (PAL), 24Hz (film)
If we point-sample time, objects have a jerky, strobbed look, e.g., sampling at ¼ rotation
Fast moving objects move large distances between frames

Temporal Aliasing

Film automatically does temporal anti-aliasing

Photographic film integrates over the exposure time
This shows up as motion blur in the photographs

Motion Blur

Temporal Aliasing

To avoid temporal aliasing we need to average over time also

Sample objects temporally

Cast multiple rays from eye through the same point in each pixel
Each of these rays intersects the scene at a different time: $ r(\vec{o}, \vec{d}, time) $
Average out results

The result is still-frame motion blur and smooth animation

Motion Blur

t_0

t_1

t_2

Shadows

Darkness caused when part or all of the illumination from a light source is blocked by an occluder (shadow caster)

Shadows

Point light sources give unrealistic hard shadows

Light sources that extend over an area (area light sources) cast soft-edged shadows

Some points see all the light: fully illuminated
Some points see none of the light source: the umbra
Some points see part of the light source: the penumbra

Hard and Soft Shadows

Shadows

Cast multiple shadow rays from surface, distributed across the surface of the light: each ray to a different point on the light source

Inside the umbra, no shadow rays get through to light
Inside the penumbra, some shadow rays get through and some parts of the are light is visible
Outside the penumbra, all shadow rays get through

Hard and Soft Shadows

Ray Tracing Area Light Source to Create Soft Shadows

Shadows

At each point, sum the contributions of shadow rays from that point to find the strength of shadow:

\[\frac{hits\cdot 100\%}{rays} = illuminated\%\]

Hard and Soft Shadows

Ray Tracing Area Light Source to Create Soft Shadows

20% illuminated

Shadows

Anti-aliasing:

Break a pixel into a grid of sub-pixels and distribute rays over the sub-pixels

Soft-shadows:

Break an area light into a grid of $N = n\times n$ point lights, each with $\frac{1}{N}$ the intensity of the base light
Sample the light, not the pixel
Each primary ray generates multiple shadow rays per light source

Sampling Area Light

Shadows

Stochastic Area-Light Sampling

Vec3f CShaderPhong::shade(const Ray& ray) {
    ...
    Ray shadow(ray.hitPoint());                              // shadow ray
    for (auto& pLight : m_scene.getLights()) {               // for each light source
        Vec3f L = Vec3f::all(0);                             // incoming randiance

        // get direction to light, and intensity
        Vec3f radiance = pLight->illuminate(shadow);         // shadow.dir is updated
        if (radiance && !m_scene.if_intersect(shadow)) {     // if not occluded
            // ------ diffuse ------
            if (m_kd > 0) {
                float cosLightNormal = shadow.dir.dot(n);
                if (cosLightNormal > 0)
                   L += m_kd * cosLightNormal * color.mul(radiance);
            }
            // ------ specular ------
            if (m_ks > 0) {
                float cosLightReflect = shadow.dir.dot(reflected.dir);
                if (cosLightReflect > 0)
                    L += m_ks * powf(cosLightReflect, m_ke) * specularColor.mul(radiance);
            }
        }
        res += L;
    }
    ...
}

Vec3f CLightOmni::illuminate(Ray& ray) {
    // ray towards point light position
    ray.dir	= m_org - ray.org;
    ray.t	= norm(ray.dir);
    ray.dir = normalize(ray.dir);
    ray.hit = nullptr;
    double attenuation = 1 / (ray.t * ray.t);
    return attenuation * m_intensity;
}

See Source Code

A point light source may be sampled with only one sample

Shadows

Stochastic Area-Light Sampling

Vec3f CShaderPhong::shade(const Ray& ray) {
    ...
    Ray shadow(ray.hitPoint());                              // shadow ray
    for (auto& pLight : m_scene.getLights()) {               // for each light source
        Vec3f L = Vec3f::all(0);                             // incoming randiance
        size_t nSamples = pLight->getNumSamples();           // number of samples
        for (size_t s = 0; s < nSamples; s++) {
            // get direction to light, and intensity
            Vec3f radiance = pLight->illuminate(shadow);     // shadow.dir is updated
            if (radiance && !m_scene.if_intersect(shadow)) { // if not occluded
                // ------ diffuse ------
                if (m_kd > 0) {
                    float cosLightNormal = shadow.dir.dot(n);
                    if (cosLightNormal > 0)
                       L += m_kd * cosLightNormal * color.mul(radiance);
                }
                // ------ specular ------
                if (m_ks > 0) {
                    float cosLightReflect = shadow.dir.dot(reflected.dir);
                    if (cosLightReflect > 0)
                        L += m_ks * powf(cosLightReflect, m_ke) * specularColor.mul(radiance);
                }
            }
        }
        res += L / nSamples                                   // average the resulting radiance
    }
    ...
}

Vec3f CLightOmni::illuminate(Ray& ray) {
    Vec2f sample = m_pSampler->getNextSample();
    Vec3f org = m_org + sample.val[0] * m_edge1 + sample.val[1] * m_edge2;

    setOrigin(org);
    Vec3f res = CLightOmni::illuminate(ray);

    double cosN = -ray.dir.dot(m_normal) / ray.t;
    if (cosN > 0) return m_area * cosN * res.value();
    else          return Vec3f::all(0);
}

See Source Code

A point light source may be sampled with only one sample

Whereas area light source need multiple samples

Shadows

Stochastic Area-Light Sampling

Vec3f CShaderPhong::shade(const Ray& ray) {
    ...
    Ray shadow(ray.hitPoint());                              // shadow ray
    for (auto& pLight : m_scene.getLights()) {               // for each light source
        Vec3f L = Vec3f::all(0);                             // incoming randiance
        size_t nSamples = pLight->getNumSamples();           // number of samples
        for (size_t s = 0; s < nSamples; s++) {
            // get direction to light, and intensity
            Vec3f radiance = pLight->illuminate(shadow);     // shadow.dir is updated
            if (radiance && !m_scene.if_intersect(shadow)) { // if not occluded
                // ------ diffuse ------
                if (m_kd > 0) {
                    float cosLightNormal = shadow.dir.dot(n);
                    if (cosLightNormal > 0)
                       L += m_kd * cosLightNormal * color.mul(radiance);
                }
                // ------ specular ------
                if (m_ks > 0) {
                    float cosLightReflect = shadow.dir.dot(reflected.dir);
                    if (cosLightReflect > 0)
                        L += m_ks * powf(cosLightReflect, m_ke) * specularColor.mul(radiance);
                }
            }
        }
        res += L / nSamples                                   // average the resulting radiance
    }
    ...
}

Vec3f CLightArea::illuminate(Ray& ray) {
    Vec2f sample = m_pSampler->getNextSample();
    Vec3f org = m_org + sample.val[0] * m_edge1 + sample.val[1] * m_edge2;

    setOrigin(org);
    Vec3f res = CLightOmni::illuminate(ray);

    double cosN = -ray.dir.dot(m_normal) / ray.t;
    if (cosN > 0) return m_area * cosN * res.value();
    else          return Vec3f::all(0);
}

See Source Code

A point light source may be sampled with only one sample

Whereas area light source need multiple samples

Shadows

Stochastic Area-Light Sampling

Area light sampling is another way to describe the standard rendering equation:

$$ L_o(p,\omega_o) = L_e(p,\omega_o) + \int_{H^2(\vec{n})} f_r(p,\omega_o,\omega_i)L_i(p,\omega_i)\cos\theta_i d\omega_i$$

by integrating not over the hemisphere, but over the area of the light source:

$$ L_o(p,\omega_o) = L_e(p,\omega_o) + \int_{t\in S} f_r(p,\omega_o,\omega_i)L_o(p,-\omega_i) \frac{\cos{\theta_i}\cos{\theta_o}}{(p-t)^2} dA_y $$

integration over a hemisphere

\omega_i

-\omega_i

integration over the light source

\omega_i

-\omega_i

t_0

t_1

\theta_0

\theta_1

Shadows

Sample Distribution

Uniform distribution gives rise to sharp transitions / patterns inside penumbra

Primary rays: 4 / pixel

Shadow rays: 1 / light (centered)

Overall: 4 rays / pixel

Primary rays: 4 / pixel

Shadow rays: 16 / light (uniform)
Overall: 64 rays / pixel

Shadows

Stochastic Sampling of Area Light

Area light represented as a rectangle in 3D, each ray-object intersection samples the area-light at random:

\[\vec{r} = \vec{c} + \xi_1\vec{a} + \xi_2\vec{b},\]

where $\xi_1$ and $\xi_2$ are random variables

Stratified sampling of the area light with samples spaced uniformly plus a small perturbation: \[\left\{\vec{r}(i, j),0\leq i,j\leq n-1\right\}\]

\vec{r}(0,0)

Shadows

Sampling Rate

As with stochastic super-sampling for anti-aliasing, light sampling rate must be high, otherwise high-frequency noise becomes visible

Primary rays: 4 / pixel

Shadow rays: 1 / light

Overall: 4 rays / pixel

Primary rays: 4 / pixel

Shadow rays: 16 / light Overall: 64 rays / pixel

Shadows

Samples Distribution

The total number of samples is the same in both cases, but because the version with 16 shadow samples per 1 image sample is able to use an Stratified Sampling, all of the shadow samples in a pixel’s area are well distributed, while in the first image the implementation here has no way to prevent them from being poorly distributed.

Primary rays: 16 / pixel

Shadow rays: 1 / light

Overall: 16 rays / pixel

Primary rays: 1 / pixel

Shadow rays: 16 / light Overall: 16 rays / pixel

Shadows

Samples Distribution

Primary rays: 16 / pixel

Shadow rays: 1 / light

Overall: 16 rays / pixel

Primary rays: 1 / pixel

Shadow rays: 16 / light Overall: 16 rays / pixel

Transformations between Distributions

Sampling a Unit Disc

4096 random samples

Random Sampling

r=\xi_1 \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Transformations between Distributions

Sampling a Unit Disc

4096 random samples

r=\xi_1 \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Random Sampling

Transformations between Distributions

Sampling a Unit Disc

4096 random samples

r=\sqrt{\xi_1} \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Random Sampling

Transformations between Distributions

Sampling a Unit Disc

4096 stratified samples

r=\sqrt{\xi_1} \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Stratified Sampling

Transformations between Distributions

Sampling a Unit Disc

4096 uniform samples

4096 stratified samples

Stratified Sampling

r=\sqrt{\xi_1} \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Transformations between Distributions

Sampling a Unit Disc

4096 uniform samples

4096 stratified samples

Concentric Sampling

r=\xi_1 \\ \theta=\frac{\xi_2}{\xi_1}\frac{\pi}{4} \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

Transformations between Distributions

Sampling a Hemisphere

Disc:

Sphere:

Uniform Sampling

r=\sqrt{\xi_1} \\ \theta=2\pi\xi_2 \\ x=r\cos{\theta} \\ y=r\sin{\theta} \\

\phi=\arccos{\xi_1} \\ \theta=2\pi\xi_2 \\ x=\sin{\phi}\cos{\theta} \\ y=\sin{\phi}\sin{\theta} \\ z=\cos{\phi}

Transformations between Distributions

Sampling a Hemisphere

Cosine-Weighted Sampling

To sample direction vectors from a cosine-weighted distribution, uniformly sample points on the unit disk and project them up to the unit hemisphere.

Vec3f CSampler::cosineSampleHemisphere(const Vec2f& sample) {
    Vec2f s = concentricSampleDisk(sample);
    float z = sqrtf(max(0.0f, 1.0f - s[0] * s[0] - s[1] * s[1]));
    return Vec3f(s[0], s[1], z);
}

Depth of Field

Aperture

Lens with big aperture

optical axis

Focus Plane

sensor

DoF

The Circle of Confusion (CoC) is the maximum size that a blur spot, on the image captured by the camera sensor, will be seen as a point in the final image.

Object is considered in focus if on an 8×10 print viewed at a distance of 10”, diameter of CoC ≤ 0.01” (1930’s standard!)

CoC

Depth of Field

Aperture

Lens with small aperture

optical axis

Focus Plane

sensor

DoF

The Circle of Confusion (CoC) is the maximum size that a blur spot, on the image captured by the camera sensor, will be seen as a point in the final image.

Object is considered in focus if on an 8×10 print viewed at a distance of 10”, diameter of CoC ≤ 0.01” (1930’s standard!)

CoC

Depth of Field

Aperture

Depth of Field

Real cameras have lenses with focal lengths

Only one plane is truly in focus
Points away from the focus project as circle of confusion
The further away from the focus the larger the circle

The range of distances that appear in focus is the depth of field

smaller apertures (larger f-numbers) result in greater depth of field

Depth of field can be simulated by distributing primary rays through different parts of a lens assembly

Depth of Field

All rays origin from point $\vec{B}$
Pixel $p_1$, which corresponds to an object point $\vec{D}$ in the focus plane uses ray $\vec{D} - \vec{B}$
Pixel $p_2$, for object $\vec{E}$ behind (or in front of) the focus plane, uses ray $\vec{E} - \vec{B}$

Implementation

Standard ray tracing

optical axis

focus plane

image plane

CoC

lens

\vec{B}

\vec{D}

p_2

\vec{E}

p_1

Depth of Field

Rays origin from the lens plane (e.g. points $\vec{A}$, $\vec{B}$ and $\vec{C}$ )
Pixel $p_1$, which corresponds to an object point $\vec{D}$ in the focus plane uses rays $\vec{p_1} - \vec{A}$, $\vec{p_1} - \vec{B}$ and $\vec{p_1} - \vec{C}$
Pixel $p_2$, for object $\vec{E}$ behind (or in front of) the focus plane, averages rays $\vec{p_2} - \vec{A}$, $\vec{p_2} - \vec{B}$ and $\vec{p_2} - \vec{C}$
To simulate more accurately, first refract primary rays through lens

Implementation

Stochastic ray tracing

optical axis

focus plane

image plane

CoC

lens

\vec{B}

p_2

\vec{E}

\vec{A}

\vec{C}

\vec{D}

p_1

Depth of Field

Or simply select eye positions randomly from a square region

Implementation

Glossy Reflections

Ray tracing simulates perfect specular reflection, true only for perfect mirrors and chrome surfaces

Most surfaces are imperfect specular reflectors:

Surface microfacets perturb direction of reflected rays
Reflect rays in a cone around perfect reflection direction
Phong specular lighting tries to fake this with the $k_e$ exponent

Glossy Reflections

For each ray-object intersection

Instead of shooting one ray in the perfect specular reflection (mirror) direction
stochastically sample multiple rays within the cone about the specular angle
Strength of reflection drops off rapidly away from the specular angle
Probability of sampling that direction should fall off similarly

Implementation

perfect mirror

polished surface

Glossy Reflections

Instead of mirror images:

Highlights can be soft
Blurred reflections of objects

Nearby objects reflect more clearly because distribution still narrow

Farther objects reflect more blurry because distribution has spread

Glossy Reflections

Scene rendered with a relatively ineffective sampler (left) and a carefully designed sampler (right), using the same number of samples for each. The improvement in image quality, ranging from the edges of the highlights to the quality of the glossy reflections, is noticeable.

Primary rays: 4 / pixel

Shadow rays: 16 / light
Reflection rays: 16 / point (random)

Primary rays: 4 / pixel

Shadow rays: 16 / light
Reflection rays: 16 / point (stratified / cosine sampling)

Glossy Reflections

Primary rays: 4 / pixel

Shadow rays: 16 / light
Reflection rays: 16 / point (random)

Primary rays: 4 / pixel

Shadow rays: 16 / light
Reflection rays: 16 / point (stratified / cosine sampling)

Translucency

Similar, but for refraction

Instead of distributing rays around the reflection ray, distribute them around the refracted ray

Transparency

Translucency

Stochastic Ray Tracing

By Dr. Sergey Kosov

Stochastic Ray Tracing

Lesson 7: www.OpenRT.org

1,197

Dr. Sergey Kosov

University Lecturer and Entrepreneur

project-10.de/Kosov

Lesson 7

Stochastic Ray Tracing

Contents

Introduction

Why does ray tracing look obviously computer generated ?

Crisp images are too "perfect":

Introduction

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Introduction

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Multiple primary rays through a pixel

Introduction

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Multiple primary rays through a pixel

Multiple shadow rays to sample an area light

Introduction

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Multiple primary rays through a pixel

Multiple shadow rays to sample an area light

Multiple reflection rays

Introduction

Distributed (Stochastic) ray tracing

Replace each single ray with a distribution of multiple rays

Multiple primary rays through a pixel

Multiple shadow rays to sample an area light

Multiple reflection rays

Multiple refraction rays

Supersampling

Supersampling

Supersampling

Super-sampling does not eliminate aliasing, it simply pushes it to higher frequencies

Other than regular, fixed sampling pattern, sampling can also be stochastic (a.k.a., random, probabilistic, or Monte Carlo)

Non-fixed Sampling Patterns

Supersampling

Stochastic Supersampling

Supersampling

Stochastic Supersampling

Supersampling

Stratified Sampling

Supersampling

1 ray per pixel

Stratified Sampling

Supersampling

16 rays per pixel (uniform)

Stratified Sampling

Supersampling

16 rays per pixel (jittered)

Stratified Sampling

Supersampling

Motion Blur

Temporal Aliasing

Temporal Aliasing

Film automatically does temporal anti-aliasing

Motion Blur

Temporal Aliasing

To avoid temporal aliasing we need to average over time also

Motion Blur

Shadows

Shadows

Hard and Soft Shadows

Shadows

Hard and Soft Shadows

Ray Tracing Area Light Source to Create Soft Shadows

Shadows

Hard and Soft Shadows

Ray Tracing Area Light Source to Create Soft Shadows

Shadows

Anti-aliasing:

Soft-shadows:

Sampling Area Light

Shadows

Stochastic Area-Light Sampling

Shadows

Stochastic Area-Light Sampling

Shadows

Stochastic Area-Light Sampling

Super-sampling does not eliminate aliasing,
it simply pushes it to higher frequencies

Other than regular, fixed sampling pattern, sampling can also be stochastic
(a.k.a., random, probabilistic, or Monte Carlo)