Topics Refs
This document is updated frequently, so remember to refresh your browser.

Thursday, December 8
  • Presentations
Tuesday, December 6
  • Presentations
Thursday, December 2
  • No class - project prep
Tuesday, November 29
  • Perlin noise
  • Code examples:
Thursday, November 17
  • Bump mapping
  • Code examples:
Note: you will need the updated version of cs336util.js in which the getModelData function also computes tangent coordinates. You will also need one or more normal maps to try out (see the examples/normalmap directory and edit lines ~160-165 as of BetterBump.js as needed).
Tuesday, November 15
  • Exam 2
Thursday, November 10
  • Shadow mapping
  • Exam review
Tuesday, November 8
  • Framebuffer objects
  • Code examples:
Thursday, November 3 (Week 11)
  • Introduction to three.js
  • The base type THREE.Object3D encapsulates a position, rotation, and scale; maintains a list of child objects
  • Rotation options (remember angles are always in radian measure)
    • To get intrinsic intrinsic rotations as for CS336Object, use rotateX(), rotateY(), rotateZ(), rotateOnAxis()
    • To get an extrinsic rotation (equivalent to left-multiplying the rotation matrix), use
            var q = new THREE.Quaternion().setFromAxisAngle(axis, angle);
  • Basic abstractions: You combine a geometry (vertex attributes) and a material (surface properties) to make a mesh, which is a subtype of Object3D.
  • Objects can be added to a scene, and can be added as children of other objects
  • lights can be added to a scene (lights are also subtypes of Object3D)
  • The render operation of a renderer takes in a scene and camera (which is a subtype of Object3D)
  • A base Object3D can serve as a dummy object in a hierarchy
  • How it works: behind the scenes, three.js dynamically generates, loads, and compiles the shader source code needed to implement the geometry and material
  • Can also incorporate your own shader code via ShaderMaterial or RawShaderMaterial
  • Take a look at the first couple of chapters of the book by Tony Parisi (see the Resources page)
  • Code examples:
Tuesday, November 1 (Week 11)
  • Cube maps - texture "coordinate" is a vector pointing to a spot on one of six images arranged in a cube. (Does not need to be normalized.)
  • A few things we can do:
  • 1) Use it as a texture for a model without its own texture coordinates
    • In vertex shader, just treat the vertex position as a vector pointing out from the origin, and pass the xyz value to fragment shader.
    • Optionally, choose a different center point for the vector
    • Fragment shader has a uniform of type samplerCube (containing the texture unit number)
    • Sample from the cube map using GLSL function textureCube
  • 2) Use it as a skybox
    • Render the cube texture on a big stationary cube centered at the origin
  • 3) Use it as an environment map (reflection and refraction)
    • Sample from the cube map using a vector R that is the reflection (or refraction) of the view vector at each vertex
    • Note: the R vector needs to be in world coordinates, so the vector E towards the view point and the normal vector N need to be in world coordinates too (this means using a normal matrix that is the inverse transpose of the model matrix, not the view*model matrix)
    • Once you have E and N in world coordinates, you can calculate R = reflect(-E, N)
  • Code examples:
Thursday, October 27 (Week 10)
  • Texture mapping, con't
  • You can use the sampled value from texture memory in many different ways, e.g.
    • replace existing surface color
    • blend with surface color
    • modulate some aspect of surface properties, e.g. specular reflectivity
    • just use texture coordinates to procedurally generate a surface effect
  • (Edit the lines labeled (1), (2), and (3) in the fragment shader for LightingWithTextureAndModel.html)
  • Texture sampling parameters
  • Magnification - one texel to many fragments
    • Causes pixelation
    • NEAREST ("box" filter - single sample) or LINEAR ("tent" filter - average of four samples)
  • Minification - many texels to one fragment
    • Causes artifacts when there is a lot of detail in the texels
    • NEAREST (single sample) or LINEAR (four samples)
    • Mipmaps - pre-calculate an average of a region of texels
    • MIPMAP_NEAREST or MIPMAP_LINEAR (interpolate between values obtained from two mip levels)
    • Anisotropic filtering - by taking more samples in one direction, can use a more detailed mip level
  • To review the basics of texture mapping, see the section "Pasting an image onto a rectangle" in Chapter 5 of the teal book
  • Code examples:
Tuesday, October 25 (Week 10)
  • Introduction to texture mapping
    • Modeler associates texture coordinates (s, t), aka (u, v), with each vertex
    • These coordinates are interpolated as varying variables
    • Fragment shader uses them to sample from a location in a 2D image
    • Sampled value can be used as surface color, or something else...
  • Setting up (see loadImagePromise and createAndLoadTexture in cs336util.js):
    1. Generate a handle for a texture
    2. Choose an active "texture unit" to use during initialization (default is 0, usually ok)
    3. Bind texture to the binding point TEXTURE_2D (or TEXTURE_CUBE)
    4. Load data (image)
    5. Set wrapping and sampling parameters
  • Rendering:
    1. In shaders:
      1. Declare attribute variable for texture coordinates
      2. Declare varying variable for texture coordinates
      3. In fragment shader, declare uniform variable of type sampler2D (or samplerCube)
      4. Use texture2D function (or textureCube to sample a value from texture based on the interpolated texture coordinates
      5. Use the resulting vec4 as a color, or in some other way
    2. In JS:
      1. Choose an active "texture unit" under which to bind the texture
      2. Bind texture to binding point
      3. Set uniform variable for sampler - value to pass is the texture unit number
  • Code examples:
(Try experimenting with the choice of texture coordinates, or the image, at the top of the file)
Thursday, October 20 (Week 9)
  • Hierarchical objects - defining "child" objects relative to a "parent" object's frame
  • The problem of defining child objects relative to a frame that has been scaled
  • In these examples, if world frame is F, and parent object's vertices are FTRSc, we want child vertices to be rendered relative to frame FTR, not FTRS
  • In traditional OpenGL, programmer would explicitly manage a matrix stack
  • Using recursion
    • Each CS336Object has an array of child objects
    • When rendering an object, it recursively renders its child objects, passing its own matrix to children to serve as the childrens' "world"
  • Dummy objects do not actually draw anything, but serve as the world for one or more child objects
  • Code examples:
Tuesday, October 18 (Week 9)
  • Shader implementation of light and material properties in Lighting3.js
  • Defining multiple lights
    • array sizes and loop bounds in GLSL have to be compile-time constants
    • options for passing array data to shaders
  • Using CS336Object to define a movable camera
    • Aside - how inheritance works in Javascript
    • Add methods for extracting the view matrix and projection matrix
  • Transforming normal vectors with the view * model will not generally work!
    • Instead, use the normal matrix, which is the inverse transpose of the view * model matrix (taking just the upper left 3 x 3 submatrix)
  • The OBJ file format for models
    • The three.js OBJ loader is in three.js-master/examples/js/loaders. You can also find a copy in examples/three/OBJLoader.js.
    • To load a local file, you'll need to access the page using a local server (see the Resources page)
    • The JS runtime loads resources asynchronously: Using async/await (See loadOBJPromise in cs336util.js)
  • Code examples:
(See the comments marked "***" to see how to add in the camera and camera controls. This example depends on the implementation of CS336Object.js and Camera.js that are in the hierarchy/ directory.)
NOTE: You will need to load the page from a local server and have the obj file in a sibling directory models/ for this to work. You will also need examples/three/OBJLoader.js. See the Resources page for suggestions on running a local server. Edit the top of the js file to change the model, and edit the main method to select face normals or vertex normals.
You can find OBJ files for the dragon model (75MB) and the hairball model (236MB) at (Note that our code is not able to handle the multi-object OBJ files or material properties.)
Thursday, October 11 (Week 8)
  • Recap of lighting and shading: The Phong or ADS lighting model (reflection model)
    • Contribution from each light source is: $$k_aL_a + k_dL_d\cos(\theta) + k_sL_s(\cos(\phi))^t$$
    • where $\vec{L} \cdot \vec{N} = \cos(\theta)$ and $\vec{R} \cdot \vec{V} = \cos(\phi)$
    • Constants above are all vector quantities with a red, green and blue component
  • Shading techniques
    • Gouraud shading: use lighting model to calculate color at vertices, interpolate color to fragments
    • Phong shading: calculate vectors L, N, and V at vertices and interpolate, then use lighting model to calculate color per-fragment
  • Issues with Gouraud shading
    • Edges are still visible, since linear interpolation of vertex colors across a face is not completely realistic (effect is exaggerated by Mach banding)
    • Severe artfacts when specular component is added in (specular highlight is either missed or exaggerated)
  • Filling out the lighting model
    • Material properties: 9 numbers (ambient, diffuse, specular reflectivity for red, green, and blue), plus one additional number for the specular exponent
    • Light properties: 9 numbers (ambient, diffuse, specular intensity for red, green, and blue)
    • Can be passed to shader as two 3 x 3 matrices plus the exponent
  • The L, N, V vectors need to be normalized after interpolation
  • Code examples:
(Experiment with the matrices representing light and material properties. Also see the Lighting2 examples from last time for demonstration of the differences between Gouraud shading and Phong shading)
Tuesday, October 11 (Week 8)
  • Exam
Thursday, October 6 (Week 7)
  • Exam review
Tuesday, October 4 (Week 7)
  • Introduction to lighting and normal vectors
  • Lambert's law: for a perfectly diffuse surface, reflected illumination at a vertex is proportional to $\cos(\theta)$, where $\theta$ is the angle between the light direction and the surface direction
    • Not affected by the location of the viewer
    • $\cos(\theta) = \vec{L} \cdot \vec{N}$, where $\vec{L}$ is a unit vector in the direction of the light and $\vec{N}$ is a unit vector perpendicular to the surface (known as a normal vector for the surface)
  • The normal vector is an additional attribute for each vertex
  • Programming diffuse lighting in the vertex shader
  • The illusion of the surface shape is determined by the direction of the normal vectors
    • Face normals - perpendicular to each triangular face
    • Vertex normals - usually obtained by averaging the face normals for adjacent triangles
  • For dot products, coordinates have to be relative to the same frame
    • Common convention - transform everything into eye coordinates for the lighting calculation
  • Adding in a constant to simulate ambient light
  • For a partially reflective surface, the specular reflection depends on the angle $\phi$ between the direction $\vec{V}$ to the viewer (camera) and the reflected direction $\vec{R}$ of the light source
  • The Phong or ADS lighting model (reflection model) consists of
    • An ambient term that is constant
    • a diffuse term that is proportional to $\vec{L} \cdot \vec{N}$
    • a specular term that is proportional to $(\vec{R} \cdot \vec{V})^t$ where $t$ is a constant affecting the apparent "shininess", and $\vec{R} \cdot \vec{V} = \cos(\phi)$
  • Code examples:
First experiment with rotating cube:

Using a sphere model, diffuse only, calculated in vertex shader (Gouraud shading):
Diffuse, calculated in fragment shader (Phong shading):
Including specular, calculated in vertex shader (Gouraud shading):
Including specular, calculated in fragment shader (Phong shading):
(See comments at the top of the Lighting2x html files. For these four examples, the shader code is in the html file, since all of them use the same Javascript code Lighting2.js. Edit the main method to change the model, or to change from vertex normals to face normals.)

Use Lighting2c.html to experiment with various values of the specular exponent.

Thursday, September 29 (Week 6)
  • Summary of coordinate systems we have seen:
    • (designer creates ->) Model coordinates
    • (model transformation ->) World coordinates
    • (view transformation ->) Eye coordinates
    • (projection transformation ->) Clip coordinates
    • (perspective division ->) Normalized device coordinates (NDC)
    • (viewport transformation ->) Window coordinates
  • Clip coordinates are 4-dimensional homogeneous coordinates with a w-component that is not normally equal to 1 (if perspective projection is being used)
  • NDC are 3-dimensional with x, y, and z between -1 and 1
    • z is NOT linearly related to the original z-value in eye space (if perspective projection is being used).
    • In fact $z_{NDC}$ is proportional to $\frac{-1}{z_{eye}}$, so it does preserve correct depth ordering
  • Window coordinates are what we see as gl_FragCoord.xy
  • By default, the viewport transformation scales x and y to range from 0 to the framebuffer width/height
  • The viewport transformation is set by the function gl.viewport($x_0$, $y_0$ width, height)
    • maps NDC x value from [-1, 1] into [$x_0$, $x_0$ + width] in window coordinates
    • maps NDC y value from [-1, 1] into [$y_0$, $y_0$ + height]
    • (These are simple Fahrenheit-to-Celsius linear rescalings)
  • gl_FragCoord.z is (usually) rescaled to be from 0 to 1
    • this can be configured with gl.depthRange(near, far)
  • Introducing the idea of the CS336Object from hw3
    • The convention TRS: scale, then rotate, then translate
    • encapsulate the scale, rotation, and translation for an object in a scene
    • instead of calculating matrices, provide operations such as "move forward" or "turn right" or "look at"
  • Code examples:
Tuesday, September 27 (Week 6)
  • When we think about Euler angles, the most convenient convention for us is usually the ordering YXZ, known as "head-pitch-roll"
  • Spherical coordinates
    • Using head and pitch angles to describe a direction
  • Rotations about an arbitrary axis
  • One strategy:
    • Let $YX$ denote the pitch and head rotation needed to align the y-axis with the desired axis of rotation.
    • Let $\theta$ be the desired angle of rotation
    • Then $YX$ * RotateY($\theta$) * $X^{-1}Y^{-1}$ is the rotation matrix
  • Or, use the THREE.Matrix4 method makeRotationAxis (See #2 in the animation loop for RotatingCubeAxisTest.js)
  • Issues with Euler angles
    • Unnatural and confusing in many cases
    • Subject to "gimbal lock" - losing a degree of freedom when two of the axes end up aligned with each other
    • Difficult to interpolate between rotations for animation
    • You can't just linearly interpolate between rotation matrices, since the intermediate values aren't rotation matrices!
      • Example: what matrix is halfway between these two rotations, if you just average the corresponding entries? \[\begin{bmatrix}1 & 0 \\ 0 & 1 \end{bmatrix} \longrightarrow \begin{bmatrix}0 & -1 \\ 1 & 0 \end{bmatrix} \]
  • Every combination of Euler angles can be described as a single rotation with some axis and angle
    • For every orthogonal 3d matrix $M$ there is a vector $\vec{v}$ such that $M\vec{v} = \vec{v}$ (an eigenvector with eigenvalue 1)
    • Therefore $\vec{v}$ is the actual axis of rotation
  • A quaternion describes a rotation as four numbers representing an axis and an angle
    • The quaternion for angle $\theta$ with axis $\shortcvec{v_1}{v_2}{v_3}^T$ is, literally, the four numbers: \[\cvec{\sin\left(\frac{\theta}{2}\right) v_1} {\sin\left(\frac{\theta}{2}\right) v_2} {\sin\left(\frac{\theta}{2}\right) v_3} {\cos\left(\frac{\theta}{2}\right)} \]
    • The scaling by sine and cosine looks bizarre, but it's basically there to make the multiplication operations work nicely...
    • ...but we don't care; we rely on library functions to do the gory stuff
    • There are library functions to create a quaternion from a rotation matrix or axis/angle, and to create a rotation matrix from a quaternion
    • The three.js library internally converts all rotations to quaternions
    • Most importantly: Quaternions can be linearly interpolated to get smooth animation between rotations
  • Spherical linear interpolation of quaternions (slerp) using the three.js library
    • Compare the ad hoc rotation about the quaternion axis (around lines 445-455 of RotationsWithQuaternion.js) to the use of the slerp method (line 477 of RotationsWithQuaternionSlerp.js)
  • See chapter 7 of the Gortler book for a brief, graphics-centric explanation of quaternions (available online through ISU library, see the Resources page)
  • Code examples:
(Put the model into any rotation, then press the 'a' key to see it rotate back to the identity, about the quaternion axis (the magenta line) Use 'A' and 'a' to do it again.)
See function calculateAxisAngle around line 250 to see how the three.js library is being used to get the axis and angle from the quaternion.
Thursday, September 22 (Week 5)
  • A bit more about perspective, depth, and the pipeline
    • Homogeneous coordinates are what we feed in to the pipeline from the vertex shader, and are used throughout the rasterization process (see GL_example1a_homogeneous)
    • After rasterization, all coordinates are divided through by their w-component to obtain Normalized Device Coordinates (NDC)
    • This step is called perspective division
    • Normalized device coordinates are in range [-1, 1]
    • The x and y values are rescaled to the canvas width and height, respectively, to get the values we see as gl_FragCoord.xy
    • The z value is normally rescaled to the range [0, 1]
    • A perspective projection maintains the linear relationships in the x and y values, but the mapping from camera z to NDC z is highly nonlinear
  • z-fighting
    • Depth buffer has finite precision, so when z-values are close together, the roundoff errors may cause a farther surface to "bleed" through to a closer one
    • The problem is worsened by the nonlinearity of the perspective projection, especially when the near plane is close to the camera (See depth_graph.pdf)
  • Arts and crafts!
    • Use a physical object to help visualize 3D rotations
  • Any possible orientation of the coordinate axes that doesn't change the lengths or the angles between the basis vectors, is a rotation
    • If you write down the coordinates of the transformed basis vectors as columns of a matrix, that's a rotation matrix
    • A rotation matrix is always orthogonal (the columns are an orthonormal set of vectors comprising the basis for the rotated frame)
    • But it's not necessarily a rotation about one of the coordinate axes
  • Composing rotations about the three coordinate axes
  • Euler angles
    • Choose an ordering of two or three axes, such as YZY or XYZ
    • Any rotation can be obtained as a sequence of three rotations about that sequence of axes
Tuesday, September 21 (Week 5)
  • Example: consider a translation $T$ and consider a set of translated points $\mathcal{F}T\underline{c}$. When we apply another transformation $R$, we can either do
    • $\mathcal{F}RT\underline{c}$ - $R$ is applied wrt the original frame $\mathcal{F}$ ("extrinsically")
    • $(\mathcal{F}T)R\underline{c}$ - $R$ is applied wrt the transformed (local) frame $\mathcal{F}T$ ("intrinsically")
  • Try the first two radio buttons in Transformations2.html
  • The "left-of"" rule: a transformation matrix is always applied with respect to the basis immediately to its left
  • To perform a transformation $M$ wrt some other frame $\mathcal{F'} = \mathcal{F}A$, use the matrix $AMA^{-1}$
  • Spinning cube example
    • updates the model transformation once each frame, multiplying it by a one-degree rotation about one of the axes
    • Edit the animation loop (lines 255-295 of to see the difference between multiplying on the left vs multiplying on the right
  • Orthographic vs perspective projections
  • Homogeneous coordinates revisited
    • For any nonzero $w$, $\cvec{wx}{wy}{wz}{w}^T$ describes the same 3D point as $\cvec{x}{y}{z}{1}^T$
    • This representation allows us to do a perspective projection by matrix multiplication
    • Also allows the rasterizer to more efficiently do perspective-dependent calculations
    • The first three coordinates are then divided by the $w$-coordinate to recover a 3D point in "Normalized Device Coordinates" (NDC)
    • (This operation is called perspective division)
  • Deriving a basic perspective matrix using similar triangles: if the center of projection is at the camera position, and the projection plane is $n$ units in front of the camera at $z = -n$, then the projected values $x$ and $y$ values are $x' = -\frac{xn}{z}$ and $y' = -\frac{yn}{z}$. Note: \[\begin{bmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & \frac{-1}{n} & 0 \end{bmatrix} \begin{bmatrix}x \\ y \\ z \\ 1 \end{bmatrix} = \begin{bmatrix}x \\ y \\ z \\ -\frac{z}{n} \end{bmatrix} \mbox{which is the same point as} \begin{bmatrix} -\frac{xn}{z} \\ -\frac{yn}{z} \\ -n \\ 1 \end{bmatrix} \]
  • Basic perspective matrix above works great for $x$ and $y$, but loses depth information since all $z$-values are collapsed to $-n$.
  • The standard OpenGL perspective matrix performs the same transformation on $x$ and $y$, but uses some mathematical trickery to keep some information about $z$ too
    • The THREE.Matrix4 function makePerspective specifies the viewing region as a rectangular frustum (chopped off rectangular pyramid) using left, right, top, bottom to describe dimensions of the near plane
    • See also the helper function createPerspectiveMatrix (in cs336util.js), which specifies the viewing region using a field of view angle plus an aspect ratio
  • The section "Specifying the Visible Range (Pyramid)" in teal book ch. 7 illustrates the perspective projection
  • This excellent page by Song Ho Anh goes over the derivation of the perspective matrix (and has some illustrations of the "viewing frustum")
  • Experiment with perspective matrices in the rotating cube example (lines 90 - 110)
  • Code examples:
(You can see the basic perspective matrix created around line 66 of the .js file. Notice we have lost the depth information, since all z-values are -4.)
Thursday, September 15 (Week 4)
  • In almost all cases, we think of all vertices as being transformed by three matrices, called the model, view, and projection (i.e. projection * view * model)
    • The model transformation takes model coordinates to world coordinates (i.e. places an instance of a model in the scene)
    • view transformation takes world coordinates to eye (camera) coordinates
    • projection transformation takes eye coordinates to clip coordinates (determines the visible region of the scene)
  • Last time: if point $\widetilde{p}$ has coordinates $\underline{c}$ w.r.t. a frame $\mathcal{F}$, and $M$ transforms $\mathcal{F}$ to a new frame $\mathcal{F}' = \mathcal{F}M$, then $\widetilde{p}$ has coordinates $M^{-1}\underline{c}$ w.r.t. $\mathcal{F}'$.
  • This is how we define a camera or view matrix: If $M$ is the transformation to the camera frame, then multiplying by $M^{-1}$ gives you the coordinates of everything in the scene, relative to the camera.
  • The idea of the projection matrix is to choose what region gets mapped into clip space. By default, this is just a cube centered at the camera's origin that goes from -1 to 1 in all three dimensions.
    • Specified by six "clipping planes" defining a rectangular region relative to the camera frame
    • Deriving an "orthographic" projection matrix: ranges [left, right], [bottom, top], and [-near, -far] are all scaled into [-1, 1] using three Fahrenheit to Celsius conversions
    • An orthographic projection does not account for perspective
    • See comment #8 in DepthWithView.js, about line 144
    • See this derivation by Song Ho Anh
  • Digression:
    • With an orthonormal basis, we can calculate the dot product of vectors from their coordinates (like in a physics book!)
    • If $\mathcal{B}$ is an orthonormal basis, $\vec{u}$ and $\vec{u}$ have coordinates $\underline{c}$ and $\underline{d}$, respectively, w.r.t. $\mathcal{B}$, then $\vec{u} \cdot \vec{v}$ = $\underline{c}^T\underline{d} = c_1d_1 + c_2d_2 + c_3d_3$
    • When $\vec{u}$ and $\vec{v}$ are unit vectors, $\vec{u} \cdot \vec{v}$ is the cosine of the angle between them
    • Thus $\vec{u} \cdot \vec{v}$ is 1 when $\vec{u}$ and $\vec{v}$ are parallel, and is 0 when they are orthogonal
    • A matrix is orthonormal (sometimes called orthogonal) if its columns represent unit-length vectors that are orthogonal to each other.
    • Key point: If $A$ is orthogonal, then $A^{-1} = A^T$ (the inverse is just the transpose)
  • The lookAt matrix: define the camera frame by specifying:
    • $\widetilde{\rm{eye}}$ - where is the camera?
    • $\widetilde{\rm{at}}$ - what's it pointed at?
    • $\vec{up}$ - which way is up?
    • Calculate basis $\shortcvec{\vec{x}}{\vec{y}}{\vec{z}}$ for camera with two cross products:
      • $\vec{z} = \widetilde{\rm{eye}} - \widetilde{\rm{at}}$, normalized
      • $\vec{x} = \vec{up} \times \vec{z}$, normalized
      • $\vec{y} = \vec{z} \times \vec{x}$
      • $R$ = matrix whose columns are the coordinates of $\vec{x},\vec{y},\vec{z}$. $R$ is orthogonal, so $R^{-1} = R^T$
      • $T$ = Translate($\widetilde{\rm{eye}})$, so $T^{-1}$ = Translate($-\widetilde{\rm{eye}}$)
      • view matrix is inverse of $TR$, which is $R^TT^{-1}$, i.e., the transpose of $R$ times Translate($-\widetilde{\rm{eye}}$)
    • See comment #9 in DepthWithView.js, about line 154
  • Code examples: Work through the detailed comments numbered 1 through 9 - uncomment the relevant code to try each one out
Tuesday, September 13 (Week 4)
  • An affine matrix is any 4x4 matrix with $\cvec{0}{0}{0}{1}$ in bottom row
  • The product of affine matrices is an affine matrix
  • An affine matrix represents an affine transformation
  • An affine matrix $M$ can always be decomposed into a linear (or "rotational") part $R$ followed by a translational part $T$, that is, $M = TR$ where $T$ is a translation and $R$ is linear \[ \begin{bmatrix}1 & 0 & 0 & t_x \\ 0 & 1 & 0 & t_y \\ 0 & 0 & 1 & t_z \\ 0 & 0 & 0 & 1 \end{bmatrix} \begin{bmatrix}a & d & g & 0 \\ b & e & h & 0 \\ c & f & i & 0 \\ 0 & 0 & 0 & 1 \end{bmatrix} = \begin{bmatrix}a & d & g & t_x \\ b & e & h & t_y \\ c & f & i & t_z \\ 0 & 0 & 0 & 1 \end{bmatrix} \]
    • Note that $RT$ is also an affine transformation, and its linear part is $R$, but its translational part is not $T$. That is, $RT = T'R$, where $T'$ is a translation that is generally not equal to $T$.
  • Matrix inverse: $AA^{-1} = A^{-1}A = I$
  • Why $(AB)^{-1} = B^{-1}A^{-1}$
  • Not every matrix has an inverse
  • Standard transformation matrices are easy to invert, e.g. inverse of Scale(2, 3, 1) is Scale(1/2, 1/3, 1)
  • Observation: Suppose point $\widetilde{p}$ has coordinates \underline{c} with respect to frame $\mathcal{F}$, and we transform by matrix $M$ to get a point $\widetilde{p} = \mathcal{F}(M\underline{c})$. There are two ways to look at this new point:
    • $\mathcal{F}(M\underline{c})$ - new coordinates, interpreted w.r.t. existing frame $\mathcal{F}$
    • ($\mathcal{F}M)\underline{c}$ - same coordinates, interpreted w.r.t a transformed frame $\mathcal{F}M$
  • For an affine matrix $M$ and frame $\mathcal{F}$, the product $\mathcal{F}M$ is also a frame, transformed by $M$
  • A different question: suppose we want to keep the same point, but find its new coordinates w.r.t a transformed frame $\mathcal{F}M$?
  • Example: if $\underline{c}$ is the coordinate vector that takes you from Main and Duff to Steve's house, and $M$ is the transformation that takes Main and Duff to the Clocktower, then how do you get from the Clocktower to Steve's house?
    • First invert $M$ to get from Clocktower to Main & Duff
    • Then use coordinates $\underline{c}$ to get to Steve's.
    • That is, $M^{-1}\underline{c}$ gives you the coordinates of Steve's house with respect to the Clocktower
  • More generally, if point $\widetilde{p}$ has coordinates $\underline{c}$ w.r.t. a frame $\mathcal{F}$, and $\mathcal{F}' = \mathcal{F}M$, then $\widetilde{p}$ has coordinates $M^{-1}\underline{c}$ w.r.t. $\mathcal{F}'$. That is: $$\begin{align} \mathcal{F}' &= \mathcal{F}M \\ \Longrightarrow \mathcal{F}'M^{-1} &= \mathcal{F} \\ \Longrightarrow \mathcal{F}'M^{-1}\underline{c} &= \mathcal{F}\underline{c} = \widetilde{p} \end{align}$$
  • The depth buffer and gl_FragCoord.z
  • The depth buffer algorithm for hidden surface removal
        initialize all locations of depth buffer to "infinity"
        in each fragment (x, y)
            let z' = value in depth buffer at (x, y)
            if gl_FragCoord.z < z'
                set depth buffer value at (x, y) to gl_FragCoord.z
                run the fragment shader
                do nothing
  • Clip space is left-handed!
    • larger z-coordinate means "farther away"
  • The section "Correctly Handling Foreground and Background Objects" in teal book ch. 7 describes the depth buffer
  • The idea of finding the coordinates of a point with respect to some other frame has a lovely explanation in Chapter 13 ("Change of basis") of the linear algebra videos at Chapters 2, 3, and 4 are also relevant to what we have been doing lately.
  • Code examples: Experiment with compositions of transformations using the key controls:
Thursday, September 8 (Week 3)
  • Composing transformations is matrix multiplication
  • If $R$ and $S$ are matrices for two transformations and $\underline{c}$ is a coordinate vector, then \[ (SR)\underline{c} = S(R\underline{c}) \] that is, the matrix $SR$ represents the transformation that does $R$ first, then $S$
  • In general $SR \neq RS$, for example, "rotate and then scale" is almost never the same transformation as "scale and then rotate"
  • Overview of matrix operations from the three.js library
  • Recall that we can use 4d homogeneous coordinates to represent 3d points and vectors.
    • The 4th coordinate is 0 for a vector and 1 for a point
    • A linear transformation can be represented by a 4x4 matrix whose bottom row and right column are both $\cvec{0}{0}{0}{1}$ (The upper left 3x3 submatrix is the same as what we derived last time.)
  • An affine transformation is a linear transformation followed by a translation (shift)
    • in one dimension: $f(x) = mx + b$
  • A translation is represented by a matrix of the form below: \[ \begin{bmatrix}1 & 0 & 0 & t_x \\ 0 & 1 & 0 & t_y \\ 0 & 0 & 1 & t_z \\ 0 & 0 & 0 & 1 \end{bmatrix} \begin{bmatrix}x \\ y \\ z \\ 1 \end{bmatrix} = \begin{bmatrix}x + t_x \\ y + t_y \\ z + t_z \\ 1 \end{bmatrix} \]
  • A vector with length 1 is called a unit vector
  • A nonzero vector can always be divided by its length to create a unit vector with the same direction (called "normalizing" the vector)
  • The dot product between two unit vectors is the cosine of the angle between them
  • Two nonzero vectors are orthogonal (perpendicular) if their dot product is zero
  • An orthonormal basis consists of three unit vectors that are orthogonal to each other
  • A standard basis is an orthonormal basis that is right-handed
  • For any standard basis, a general rotation matrix (for rotations of angle $\theta$ about the z-axis) looks like this: \[ \begin{bmatrix}\cos(\theta) & -\sin(\theta) & 0 & 0 \\ \sin(\theta) & \cos(\theta) & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{bmatrix} \]
  • "Pseudocode" for standard matrices
    • RotateX($\theta$), RotateY($\theta$), RotateZ($\theta$)
    • Scale($s_x, s_y, s_z$)
    • Translate($t_x, t_y, t_z$)
  • See sections 3.2 - 3.5 of Gortler (ignore 3.6 on normal vectors for now)
  • The teal book section "Translate and Then Rotate" in ch. 4 goes into composing multiple transformations
  • You'll need the Matrix4 type from three.js. You can download the core library from Canvas, see link #6 on our Canvas front page. See for more info.
  • See the documentation for the three.js Matrix4 type
  • Also see chapter 4 of the Linear Algebra series by 3blue1brown:
  • Code examples (edit the main method to try different transformations):
Tuesday, September 6 (Week 3)
  • Using an index buffer to select order of vertices
  • Linear transformations
    • key feature: if you transform points that are collinear, then the resulting points are also collinear...
    • triangles are transformed to triangles
  • For a given basis $\mathcal{B} = \shortcvec{\vec{e_1}}{\vec{e_1}}{\vec{e_1}}$, a linear transformation $f$ is represented by a matrix $M$ such that if $\underline{c}$ is a coordinate vector w.r.t. $\mathcal{B}$ for some vector $\vec{u}$, then $M\underline{c}$ is the coordinate vector for the transformed vector $f(\vec{u})$
  • The columns of $M$ are just the coordinate vectors for $f(\vec{e_1}), f(\vec{e_2})$, and $f(\vec{e_3})$
  • Examples of 3d linear transformations and their matrices - 90 degree rotation, scaling
  • Using gl.uniform4fv to pass matrix data to the GPU
  • Matrix data is passed to the GPU in column-major order
  • See sections 2.3 and 3.3 of Gortler
  • In the teal book, the section "Moving, Rotating, and Scaling" in chapter 3 is a basic introduction to transformations
  • Also see chapter 3 of the Linear Algebra series by 3blue1brown:
  • Code examples (edit the main method to try different transformations):
Thursday, September 1 (Week 2)
  • In graphics, we need to deal with multiple "frames of reference" (coordinate systems)
  • Preview of typical vertex processing steps:
    • Model transformation places instance of model into world coordinates
    • View transformation transforms world into eye/camera coordinates (scene from eye/camera view point)
    • Projection transformation transforms a viewable region of scene into clip coordinates (a 2x2x2 cube representing the "viewable" triangles to be handed to the rasterizer)
  • A point is a geometric location (and may have different coordinates depending on the frame)
  • The difference between two points is a vector, representing the distance and direction from one point to another
  • A point plus a vector is a point
  • Vector operations - scaling and addition
  • Linear combinations of vectors (scale and add)
  • A 3d vector space consists of all 3D vectors with the basic operations above
  • A 3d basis for a vector space is a set of three independent vectors $\vec{b_1}, \vec{b_2}, \vec{b_3}$ such that every possible 3d vector can be written as a linear combination of $\vec{b_1}, \vec{b_2}, \vec{b_3}$
  • A 3d affine space consists of all 3d vectors and points
  • A 3d frame (coordinate system) is a basis along with a designated point called the origin
  • Matrices and matrix multiplication
  • Representing a basis as a row matrix of three vectors, $\mathcal{B} = \shortcvec{\vec{b_1}}{\vec{b_2}}{\vec{b_3}}$
  • Representing a frame as a row matrix of three vectors and a point, $\mathcal{F} = \cvec{\vec{b_1}}{\vec{b_2}}{\vec{b_3}}{\widetilde{o}}$
  • If vector $\vec{u} = u_1\vec{b_1} + u_2\vec{b_2} + u_3\vec{b_3}$, then the column matrix \[ \underline{c} = \begin{bmatrix} u_1 \\ u_2 \\ u_3 \\ 0 \end{bmatrix} = \cvec{u_1}{u_2}{u_3}{0}^T \] is a coordinate vector for $\vec{u}$ with respect to frame $\mathcal{F}$. Note that in terms of matrix multiplication, we can write \[ \mathcal{F}\underline{c} = \cvec{\vec{b_1}}{\vec{b_2}}{\vec{b_3}}{\widetilde{o}}\cvec{u_1}{u_2}{u_3}{0}^T = \vec{u} \] (where the superscript $T$ indicates the matrix transpose)
  • Likewise, if point $\widetilde{p} = u_1\vec{b_1} + u_2\vec{b_2} + u_3\vec{b_3} + \widetilde{o}$, then the column matrix $\cvec{u_1}{u_2}{u_3}{1}^T$ is a coordinate vector for $\widetilde{p}$ w.r.t. frame $\mathcal{F}$,
  • Key point: the coordinates of a vector or point depend on what basis or frame is being used, and the coordinates only have meaning with respect to a basis or frame
  • Read Sections 2.1, 2.2, and 3.1 of the book by Gortler (Foundations of 3D computer graphics). This is just a few pages and summarizes pretty much everything we talked about today. Just log into the ISU library and search for the title. The entire book is available online. You will need to log into the ISU VPN if you are off campus.
  • Review basic matrix operations if you have not seen them before
  • For nice visuals and simple explanations of vectors and linear combinations, see chapters 1 and 2 (the second and third videos) of the Linear Algebra series by 3blue1brown:
Tuesday, August 30 (Week 2)
  • Recap of basic steps in our Hello, World example
    • The purpose of an attribute variable in the vertex shader is to pass per-vertex data from CPU to GPU
    • The function vertexAttribPointer associates the vertex attribute with the data in your buffer
  • Vertex shader must always set the built-in variable gl_Position
  • Fragment shader normally sets the built-in variable gl_FragColor
  • The built-in variable gl_FragCoord available in the fragment shader
    • gl_FragCoord.x and gl_FragCoord.y represent the position of the fragment within the framebuffer, ranging from 0 up to the canvas width/height
    • (gl_FragCoord.z is the "depth" information, stored in a separate buffer called the depth buffer. It ranges from 0 to 1.)
  • Animation by updating a uniform variable in each frame
  • Using requestAnimationFrame to create animation loop
  • There are three kinds of variable modifiers in GLSL
    • attribute variables are used to pass per-vertex data from the CPU to GPU
    • uniform variables are used to pass uniform data from CPU to GPU (same value in every vertex/fragment shader instance)
    • varying variables are used to pass data from vertex shader to fragment shader (values are interpolated by the rasterizer)
  • Using functions such as
    • uniform1f - set a uniform with one float value
    • uniform4f - set a vec4 uniform with four floating point values
    • uniform4fv - - set a vec4 uniform with an array of values
    • etc...
  • See the "WebGL Reference Card" for GLSL types and functions
  • Linear interpolation!
  • Basic example: Converting Fahrenheit to Celsius
    • Suppose we have some Farenheit temperature $x$ and we want the Celsius temperature $y$. Then $$\beta = \frac{x - 32}{212 - 32}$$ tells you "how far" $x$ is along the scale from 32 to 212. We want to go the same fraction of the way along the Celsius scale from 0 to 100, i.e., $$y = 0 + \beta(100 - 0)$$ To say that the Fahrenheit and Celsius scales are "linearly related" just means, e.g., that if we are 25% of the way from 32 to 212 in Fahrenheit, we should be 25% of the way from 0 to 100 in Celsius. More generally, to map the 32-to-212 scale to any range $A$ to $B$, you have $$\begin{eqnarray} y &=& A + \beta\cdot(B - A) \\ &=& (1 - \beta)\cdot A + \beta\cdot B \\ &=& \alpha\cdot A + \beta\cdot B \end{eqnarray}$$ where $\alpha = 1 - \beta$. One way to think of this is that you have two quantities $A$ and $B$ to be "mixed", and the proportion $\beta$ tells you "how much $B$" and $\alpha$ is "how much $A$". This is nice because it generalizes to interpolating within a triangle using barycentric coordinates.
  • Code examples:
  • Note: GL_example1a is the same as GL_example1 except that all the boilerplate helper functions have been moved into ../util/cs336util.js
Thursday, August 25 (Week 1)
  • Odds and ends:
    • Basic application structure; the onload event in JS. See foo.html and foo.js (output goes to JS console).
    • Setting up a WebGL context and clearing the canvas (see GL_example0)
    • RGBA encoding of color as four floats
    • The graphics context is a state machine (function calls rely on lots of internal state)
      • The idea of binding a buffer or shader to become "the one I'm currently talking about"
  • Overview of the steps involved in our "Hello, World!" application
    • (Initialization)
      • Create context
      • Load and compile shaders
      • Create buffers
      • Bind each buffer and fill with data
    • (Each frame)
      • Bind shader
      • For each attribute...
        • Find attribute index
        • Enable the attribute
        • Bind a buffer with the attribute data
        • Set attribute pointer to the buffer
      • Set uniform variables, if any
      • Draw, specifying primitive type
  • Options for primitives: TRIANGLES, LINES, LINE_STRIP, etc.
  • See ListExample.html for a demo
  • Chapters 2 and 3 of the teal book provide a detailed and careful overview of the steps described above.
  • Also highly recommended: the first chapter of
  • Read and experiment with GL_Example1 below
    • try changing the vertices
    • try the commented-out lines in the draw() function
  • Code examples:
(You can view the associated html and javascript source in the developer tools (Ctrl-Alt-i), or just grab everything directly from the examples/intro/ directory of )
Tuesday, August 23 (Week 1)
  • Introduction
  • This is a course in 3D rendering using OpenGL, not a course in developing GUIs!
  • WebGL is a set of browser-based JavaScript bindings for OpenGL ES 2.0, which is essentially OpenGL 3.2 with the deprecated stuff and fancy features removed
  • Overview of the GPU pipeline:
    • (Model - a set of vertices organized into a "mesh" of triangles)
      • -> Vertex processing (*)
      • -> Primitive assembly (and clipping)
      • -> Rasterization
      • -> Fragment processing (*)
    • -> (Framebuffer - graphics memory mapped to an actual display window)
  • (*) Vertex and fragment processing stages are programmed via "shaders" using GLSL, the OpenGL shading language
  • Read the syllabus
  • See the Resources page for textbook information
  • Learn JavaScript (see Resources page for ideas)