csci5607/exam-2/exam2.md

---
geometry: margin=2cm
output: pdf_document
title: Exam 2
subtitle: CSCI 5607
date: \today

author: |
  | Michael Zhang
  | zhan4854@umn.edu $\cdot$ ID: 5289259
---

\renewcommand{\c}[1]{\textcolor{gray}{#1}}
\newcommand{\now}[1]{\textcolor{blue}{#1}}
\newcommand{\todo}[0]{\textcolor{red}{\textbf{TODO}}}

[ 1 2 3 6 8 9 ]

## Reflection and Refraction

1. \c{Consider a sphere $S$ made of solid glass ($\eta$ = 1.5) that has radius
   $r = 3$ and is centered at the location $s = (2, 2, 10)$ in a vaccum ($\eta =
   1.0$). If a ray emanating from the point $e = (0, 0, 0)$ intersects $S$ at a
   point $p = (1, 4, 8)$:}

   a. \c{(2 points) What is the angle of incidence $\theta_i$?}

   The incoming ray is in the direction $I = p - e = (1, 4, 8)$, and the normal
   at that point is $N = p - s = (1, 4, 8) - (2, 2, 10) = (1, -2, 2)$. The angle
   can be found by taking the opposite of the incoming ray $-I$ and using the
   formula $\cos \theta_i = \frac{-I \cdot N}{|I| |N|} = \frac{(-1, -4, -8)
   \cdot (1, -2, 2)}{9 \cdot 3} = \frac{-1 + 8 - 16}{27} = -\frac{1}{3}$. So the
   angle $\boxed{\theta_i = \cos^{-1}(-\frac{1}{3})}$.

   b. \c{(1 points) What is the angle of reflection $\theta_r$?}

   The angle of reflection always equals the angle of incidence, $\theta_r =
   \theta_i = \boxed{\cos^{-1}(-\frac{1}{3})}$.

   c. \c{(3 points) What is the direction of the reflected ray?}

   The reflected ray can be found by first projecting the incident ray $-I$ onto
   the normalized normal $N$, which is $v = N \times |-I|\cos(\theta_i) =
   (\frac{1}{3}, -\frac{2}{3}, \frac{2}{3}) \times 9 \times \frac{1}{3} = (-1,
   2, -2)$. Then, we know the point on N where this happened is $p' = p + v =
   (1, 4, 8) + (-1, 2, -2) = (0, 6, 6)$.

   Now, we can subtract this point from where the ray originated to know the
   direction to add in the other direction, which is still $(0, 6, 6)$ in this
   case since the ray starts at the origin. Adding this to the point $p'$ gets
   us $(0, 12, 12)$, which means a point from the origin will get reflected to
   $(0, 12, 12)$.

   Finally, subtract the point to get the final answer $(0, 12, 12) - (1, 4, 8)
   = \boxed{(-1, 8, 4)}$.

   d. \c{(3 points) What is the angle of transmission $\theta_t$?}

   Using Snell's law, we know that $\eta_i \sin \theta_i = \eta_t \sin \theta_t
   = 1.0 \times \sin(\cos^{-1}(-\frac{1}{3})) = 1.5 \times \sin(\theta_t)$. To
   find the angle $\theta_t$ we can just solve: $\theta_t =
   \sin^{-1}(\frac{2}{3} \times \sin(\cos^{-1}(-\frac{1}{3}))) \approx
   \boxed{0.6796}$ (in radians).

   e. \c{(4 points) What is the direction of the transmitted ray?}

   We can just add the angle obtained above to the inverse of the normal $n' =
   (-1, 2, -2)$ (since that will be transmitted through the material). For the
   tangent direction, we can use the vector from the ray's origin to the
   incident ray projected onto the normal, which we'll call $v = (0, 6, 6)$.

   Now we know that the relationship between the coefficients of $n'$ and $v$ is
   $\tan(\theta_t)$. We can add these to obtain a point along the transmitted
   ray (it doesn't actually matter where this occurs since we're going to
   normalize this anyway, the proportion stays consistent regardless). So using
   the unit length for the inverse normal which is $\frac{n'}{||n'||} =
   (\frac{1}{3}, -\frac{2}{3}, \frac{2}{3})$ we know that the perpendicular
   vector will be $\tan(\theta_t) \times (0, \frac{1}{\sqrt{2}},
   \frac{1}{\sqrt{2}})$.

   The final step is then adding these vectors together and subtracting the
   interersection point: $(\frac{1}{3}, -\frac{2}{3}, \frac{2}{3}) +
   \tan(\theta_t) \times (0, \frac{1}{\sqrt{2}}, \frac{1}{\sqrt{2}}) - (1, 4,
   8)$.

   \todo all the numbers are wrong

## Geometric Transformations

2. \c{(8 points) Consider the airplane model below, defined in object
   coordinates with its center at $(0, 0, 0)$, its wings aligned with the $\pm
   x$ axis, its tail pointing upwards in the $+y$ direction and its nose facing
   in the $+z$ direction. Derive a sequence of model transformation matrices
   that can be applied to the vertices of the airplane to position it in space
   at the location $p = (4, 4, 7)$, with a direction of flight $w = (2, 1, -2)$
   and the wings aligned with the direction $d = (-2, 2, -1)$.}

   The order we want is (1) do all the rotations, and then (2) translate to the
   spot we want. The rotation is done in multiple steps:

   - First we want to make sure the nose of the plane points in the correct
     direction in the $xz$ plane (rotating around the $y$ axis). The desired
     resulting direction is $(2, y, -2)$, so that means for a nose currently facing
     the $+z$ direction which is $(0, y, 1)$, we want to rotate around by around
     $-\frac{3}{4}\pi$. The transformation matrix is:

     $$
     M_1 = \begin{bmatrix}
        1 & 0 & 0 & 0 \\
        0 & 1 & 0 & 0 \\
        0 & 0 & 1 & 0 \\
        0 & 0 & 0 & 1 \\
     \end{bmatrix}
     $$

   - Then we want to rotate the plane vertically, so it's pointing in the right
     direction.

   $$
   \begin{bmatrix}
     1 & 0 & 0 & x \\
     0 & 1 & 0 & y \\
     0 & 0 & 1 & z \\
     0 & 0 & 0 & 1 \\
   \end{bmatrix}
   =
   \begin{bmatrix}
     1 & 0 & 0 & 4 \\
     0 & 1 & 0 & 4 \\
     0 & 0 & 1 & 7 \\
     0 & 0 & 0 & 1 \\
   \end{bmatrix}
   $$

   Since the direction of flight was originally $(0, 0, 1)$, we have to
   transform it to $(2, 1, -2)$.

3. \c{Consider the earth model shown below, which is defined in object
   coordinates with its center at $(0, 0, 0)$, the vertical axis through the
   north pole aligned with the direction $(0, 1, 0)$, and a horizontal plane
   through the equator that is spanned by the axes $(1, 0, 0)$ and $(0, 0, 1)$.}

   a. \c{(3 points) What model transformation matrix could you use to tilt the
   vertical axis of the globe by $23.5^\circ$ away from $(0, 1, 0)$, to achieve
   the pose shown in the image on the right?}

   You could use a 3D rotation matrix. Since the axis of rotation is the
   $z$-axis, the rotation matrix would look like:

   $$
   M =
   \begin{bmatrix}
     \cos(23.5^\circ) & \sin(23.5^\circ) & 0 & 0 \\
     -\sin(23.5^\circ) & \cos(23.5^\circ) & 0 & 0 \\
     0 & 0 & 1 & 0 \\
     0 & 0 & 0 & 1 \\
   \end{bmatrix}
   $$

   b. \c{(5 points) What series of rotation matrices could you apply to the
   globe model to make it spin about its tilted axis of rotation, as suggested
   in the image on the right?}

   One way would be to rotate it back to its normal orientation, apply the spin
   by whatever angle $\theta(t)$ it was spinning at at time $t$, and then put it
   back into its $23.5^\circ$ orientation. This rotation is done around the
   $y$-axis, so the matrix looks like:

   $$
   M' =
   M^{-1}
   \begin{bmatrix}
     \cos(\theta(t)) & 0 & \sin(\theta(t)) & 0 \\
     0 & 1 & 0 & 0 \\
     -\sin(\theta(t)) & 0 & \cos(\theta(t)) & 0 \\
     0 & 0 & 0 & 1 \\
   \end{bmatrix}
   M
   $$

   c. \c{[5 points extra credit] What series of rotation matrices could you use
   to send the tilted, spinning globe model on a circular orbit of radius $r$
   around the point $(0, 0, 0)$ within the $xz$ plane, as illustrated below?}

   In the image, the globe itself does not rotate, but I'm going to assume it
   revolves around the sun at a different angle $\phi(t)$. The solution here
   would be to

## The Camera/Viewing Transformation

4. \c{Consider the viewing transformation matrix $V$ that enables all of the
   vertices in a scene to be expressed in terms of a coordinate system in which
   the eye is located at $(0, 0, 0)$, the viewing direction ($-n$) is aligned
   with the $-z$ axis $(0, 0, -1)$, and the camera's 'up' direction (which
   controls the roll of the view) is aligned with the $y$ axis (0, 1, 0).}

   a. \c{(4 points) When the eye is located at $e = (2, 3, 5)$, the camera is
   pointing in the direction $(1, -1, -1)$, and the camera's 'up' direction is
   $(0, 1, 0)$, what are the entries in $V$?}

   - Viewing direction is $(1, -1, -1)$.
   - Normalized $n = (\frac{1}{\sqrt{3}}, -\frac{1}{\sqrt{3}}, -\frac{1}{\sqrt{3}})$.
   - $u = up \times n = (\frac{1}{\sqrt{2}}, 0, \frac{1}{\sqrt{2}})$.
   - $v = n \times u = (\frac{\sqrt{6}}{6}, \frac{2}{\sqrt{6}}, -\frac{\sqrt{6}}{6})$
   - $d_x = - (eye \cdot u) = - (2 \times \frac{1}{\sqrt{2}} + 5 \times \frac{1}{\sqrt{2}}) = -\frac{7}{\sqrt{2}}$
   - $d_y = - (eye \cdot v) = - (2 \times \frac{1}{\sqrt{6}} + 3 \times
       \frac{2}{\sqrt{6}} - 5 \times \frac{1}{\sqrt{6}}) = -\frac{3}{\sqrt{6}}$
   - $d_z = - (eye \cdot n) = - (2 \times \frac{1}{\sqrt{3}} - 3 \times
       \frac{1}{\sqrt{3}} - 5 \times \frac{1}{\sqrt{3}}) = -\frac{6}{\sqrt{3}}$

   $$
   \begin{bmatrix}
      \frac{1}{\sqrt{2}}         & 0          & \frac{1}{\sqrt{2}}         & -\frac{7}{\sqrt{2}}      \\
      \frac{1}{\sqrt{6}}  & \frac{2}{\sqrt{6}} & -\frac{1}{\sqrt{6}} &  -\frac{3}{\sqrt{6}}     \\
      -\frac{1}{\sqrt{3}} & \frac{1}{\sqrt{3}} & \frac{1}{\sqrt{3}} & -\frac{6}{\sqrt{3}}    \\
      0          & 0          & 0          &  1        \\
   \end{bmatrix}
   $$

   Also solved using a Python script:

   ```py
   def view_matrix(camera_pos, view_dir, up_dir):
     n = unit(-view_dir)
     u = unit(np.cross(up_dir, n))
     v = np.cross(n, u)
     return np.array([
       [u[0], u[1], u[2], -np.dot(camera_pos, u)],
       [v[0], v[1], v[2], -np.dot(camera_pos, v)],
       [n[0], n[1], n[2], -np.dot(camera_pos, n)],
       [0, 0, 0, 1],
     ])

   camera_pos = np.array([2, 3, 5])
   view_dir = np.array([1, -1, -1])
   up_dir = np.array([0, 1, 0])
   V = view_matrix(camera_pos, view_dir, up_dir)
   print(V)
   ```

   b. \c{(2 points) How will this matrix change if the eye moves forward in the
   direction of view? [which elements in V will stay the same? which elements
   will change and in what way?]}

   If the eye moves forward, the eye _position_ and everything that depends on it
   will change, while everything else doesn't.

   | $n$  | $u$  | $v$  | $d$       |
   | ---- | ---- | ---- | --------- |
   | same | same | same | different |

   The $n$ is the same because the viewing direction does not change.

   c. \c{(2 points) How will this matrix change if the viewing direction spins
   in the clockwise direction around the camera's 'up' direction? [which
   elements in V will stay the same? which elements will change and in what
   way?]}

   In this case, the eye _position_ stays the same, and everything else changes.

   | $n$       | $u$       | $v$       | $d$  |
   | --------- | --------- | --------- | ---- |
   | different | different | different | same |

   d. \c{(2 points) How will this matrix change if the viewing direction rotates
   directly upward, within the plane defined by the viewing and 'up' directions?
   [which elements in V will stay the same? which elements will change and in
   what way?]}

   In this case, the eye _position_ stays the same, and everything else changes.

   | $n$       | $u$       | $v$       | $d$  |
   | --------- | --------- | --------- | ---- |
   | different | different | different | same |

5. \c{Suppose a viewer located at the point $(0, 0, 0)$ is looking in the $-z$
   direction, with no roll ['up' = $(0, 1 ,0)$], towards a cube of width 2,
   centered at the point $(0, 0, -5)$, whose sides are colored: red at the plane
   $x = 1$, cyan at the plane $x = -1$, green at the plane $y = 1$, magenta at
   the plane $y = -1$, blue at the plane $z = -4$, and yellow at the plane $z =
   -6$.}

   a. \c{(1 point) What is the color of the cube face that the user sees?}

   \boxed{\textrm{Blue}}

   b. \c{(3 points) Because the eye is at the origin, looking down the $-z$ axis
   with 'up' = $(0,1,0)$, the viewing transformation matrix $V$ in this case is
   the identity $I$. What is the model matrix $M$ that you could use to rotate
   the cube so that when the image is rendered, it shows the red side of the
   cube?}

   You would have to do a combination of (1) translate to the origin, (2) rotate
   around the origin, and then (3) untranslate back. This way, the eye position
   doesn't change.

   $$
   M =
   \begin{bmatrix}
      1 & 0 & 0 & 0 \\
      0 & 1 & 0 & 0 \\
      0 & 0 & 1 & -5 \\
      0 & 0 & 0 & 1 \\
   \end{bmatrix} \cdot
   \begin{bmatrix}
      0 & 0 & -1 & 0 \\
      0 & 1 & 0 & 0 \\
      1 & 0 & 0 & 0 \\
      0 & 0 & 0 & 1 \\
   \end{bmatrix} \cdot
   \begin{bmatrix}
      1 & 0 & 0 & 0 \\
      0 & 1 & 0 & 0 \\
      0 & 0 & 1 & 5 \\
      0 & 0 & 0 & 1 \\
   \end{bmatrix}
   =
   \boxed{\begin{bmatrix}
      0 & 0 & -1 & -5 \\
      0 & 1 & 0 & 0 \\
      1 & 0 & 0 & -5 \\
      0 & 0 & 0 & 1 \\
   \end{bmatrix}}
   $$

   To verify this, testing with an example point $(1, 1, -4)$ yields:

   $$
   \begin{bmatrix}
      0 & 0 & -1 & -5 \\
      0 & 1 & 0 & 0 \\
      1 & 0 & 0 & -5 \\
      0 & 0 & 0 & 1 \\
   \end{bmatrix}
   \cdot
   \begin{bmatrix}
      1 \\ 1 \\ -4 \\ 1
   \end{bmatrix}
   =
   \begin{bmatrix}
      -1 \\ 1 \\ -4 \\ 1
   \end{bmatrix}
   $$

   c. \c{(4 points) Suppose now that you want to leave the model matrix $M$ as
   the identity. What is the viewing matrix $V$ that you would need to use to
   render an image of the scene from a re-defined camera configuration so that
   when the scene is rendered, it shows the red side of the cube? Where is the
   eye in this case and in what direction is the camera looking?}

   For this, a different eye position will have to be used. Instead of looking
   from the origin, you could view it from the red side, and then change the
   direction so it's still pointing at the cube.

   - eye is located at $(5, 0, -5)$
   - viewing direction is $(-1, 0, 0)$
   - $n = (1, 0, 0)$
   - $u = up \times n = (0, 0, -1)$
   - $v = n \times u = (0, 1, 0)$
   - $d = (-5, 0, -5)$

   The final viewing matrix is $\boxed{\begin{bmatrix}
     0 & 0 & -1 & -5 \\
     0 & 1 & 0 & 0 \\
     1 & 0 & 0 & -5 \\
     0 & 0 & 0 & 1 \\
   \end{bmatrix}}$. Turns out it's the same matrix! Wow!

## The Projection Transformation

6. \c{Consider a cube of width $2\sqrt{3}$ centered at the point $(0, 0,
   -3\sqrt{3})$, whose faces are colored light grey on the top and bottom $(y =
   \pm\sqrt{3})$, dark grey on the front and back ($z = -2\sqrt{3}$ and $z =
   -4\sqrt{3}$), red on the right $(x = \sqrt{3})$, and green on the left $(x =
   -\sqrt{3})$.}

   a. \c{Show how you could project the vertices of this cube to the plane $z =
   0$ using an orthographic parallel projection:}

   i) \c{(2 points) Where will the six vertex locations be after such a
   projection, omitting the normalization step?}

   - $[\begin{matrix}-1.732 & -1.732 & -6.928\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & -1.0 & 7.0\end{matrix}]$
   - $[\begin{matrix}-1.732 & -1.732 & -3.464\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & -1.0 & 5.0\end{matrix}]$
   - $[\begin{matrix}-1.732 & 1.732 & -6.928\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & 1.0 & 7.0\end{matrix}]$
   - $[\begin{matrix}-1.732 & 1.732 & -3.464\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & 1.0 & 5.0\end{matrix}]$
   - $[\begin{matrix}1.732 & -1.732 & -6.928\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & -1.0 & 7.0\end{matrix}]$
   - $[\begin{matrix}1.732 & -1.732 & -3.464\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & -1.0 & 5.0\end{matrix}]$
   - $[\begin{matrix}1.732 & 1.732 & -6.928\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & 1.0 & 7.0\end{matrix}]$
   - $[\begin{matrix}1.732 & 1.732 & -3.464\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & 1.0 & 5.0\end{matrix}]$

   ii) \c{(1 points) Sketch the result, being as accurate as possible and
   labeling the colors of each of the visible faces.}

   This is just a square with the dark gray side facing the camera. The other
   sides are not visible because the cube is parallel to the axis, and when you
   do an orthographic projection, those faces are lost.

   iii) \c{(2 points) Show how you could achieve this transformation using one or
   more matrix multiplication operations. Specify the matrix entries you would
   use, and, if using multiple matrices, the order in which they would be
   multiplied.}

   Actually, I got the numbers above by using the three transformation matrices
   in this Python script:

   ```py
   def ortho(points):
     left = min(map(lambda p: p[0], points))
     right = max(map(lambda p: p[0], points))
     bottom = min(map(lambda p: p[1], points))
     top = max(map(lambda p: p[1], points))
     near = min(map(lambda p: p[2], points))
     far = max(map(lambda p: p[2], points))

     step_1 = np.array([
       [1, 0, 0, 0],
       [0, 1, 0, 0],
       [0, 0, -1, 0],
       [0, 0, 0, 1],
     ])

     step_2 = np.array([
       [1, 0, 0, -(left + right) / 2.0],
       [0, 1, 0, -(bottom + top) / 2.0],
       [0, 0, 1, -(near + far) / 2.0],
       [0, 0, 0, 1],
     ])

     step_3 = np.array([
       [2.0 / (right - left), 0, 0, 0],
       [0, 2.0 / (top - bottom), 0, 0],
       [0, 0, 2.0 / (far - near), 0],
       [0, 0, 0, 1],
     ])

     M_ortho = step_3 @ step_2 @ step_1
   ```

   b. \c{Show how you could project the vertices of this cube to the plane $z =
   0$ using an oblique parallel projection in the direction $d = (1, 0,
   \sqrt{3})$:}

   i) \c{(3 points) Where will the six vertex locations be after such a
   projection, omitting the normalization step?}

   ii) \c{(2 points) Sketch the result, being as accurate as possible and
   labeling the colors of each of the visible faces.}

   iii) \c{(4 points) Show how you could achieve this transformation using one
   or more matrix multiplication operations. Specify the matrix entries you
   would use, and, if using multiple matrices, the order in which they would be
   multiplied.}

7. \c{Consider the simple scene shown in the image below, where two cubes, one
   of height 1 and one of height 2, are both resting on a horizontal groundplane
   ($y = -\frac{1}{2}$), with the smaller cube’s front face aligned with $z =
   -4$ and the larger cube’s front face aligned with $z = -7$.}

   a. \c{(5 points) Let the camera location be (0, 0, 0), looking down the $-z$
   axis, with the field of view set at $90^\circ$. Determine the points, in the
   image plane, to which each of the cube vertices will be projected and sketch
   the result to scale. Please clearly label the coordinates to avoid
   ambiguity.}

   For this part, I reimplemented the perspective rendering algorithm using
   Python.

   ```py
   def perspective_matrix(left, right, bottom, top, near, far):
     return np.array([
       [2.0 * near / (right - left), 0, (right + left) / (right - left), 0],
       [0, 2.0 * near / (top - bottom), (top + bottom) / (top - bottom), 0],
       [0, 0, -(far + near) / (far - near), -(2.0 * far * near) / (far - near)],
       [0, 0, -1, 0],
     ])

   def view_matrix(camera_pos, view_dir, up_dir):
     n = unit(-view_dir)
     u = unit(np.cross(up_dir, n))
     v = np.cross(n, u)
     return np.array([
       [u[0], u[1], u[2], -np.dot(camera_pos, u)],
       [v[0], v[1], v[2], -np.dot(camera_pos, v)],
       [n[0], n[1], n[2], -np.dot(camera_pos, n)],
       [0, 0, 0, 1],
     ])
   ```

   The perspective and view matrices are:

   $$
   PV =
   \begin{bmatrix}
      1.0 & 0.0 & 0.0 & 0.0 \\
      0.0 & 1.0 & 0.0 & 0.0 \\
      0.0 & 0.0 & -1.2222 & -2.2222 \\
      0.0 & 0.0 & -1.0 & 0.0 \\
   \end{bmatrix}
   \begin{bmatrix}
      1.0 & 0.0 & 0.0 & -0.0 \\
      0.0 & 1.0 & 0.0 & -0.0 \\
      0.0 & 0.0 & 1.0 & -0.0 \\
      0.0 & 0.0 & 0.0 & 1.0 \\
   \end{bmatrix}
   $$

   Then I ran the transformation using the data given in this particular scene:

   ```py
   def compute_view(near, vfov, hfov):
     left = -math.tan(hfov / 2.0) * near
     right = math.tan(hfov / 2.0) * near
     bottom = -math.tan(vfov / 2.0) * near
     top = math.tan(vfov / 2.0) * near
     return left, right, bottom, top

   def solve(camera_pos, angle):
     angle_radians = math.radians(angle)
     near = 1
     far = 10
     view_dir = np.array([0, 0, -1])
     up_dir = np.array([0, 1, 0])
     left, right, bottom, top = compute_view(near, angle_radians, angle_radians)
     P = perspective_matrix(left, right, bottom, top, near, far)
     V = view_matrix(camera_pos, view_dir, up_dir)
     return P @ V

   camera_pos = np.array([0, 0, 0])
   angle = 90
   m = np.around(solve(camera_pos, angle), 4)
   ```

   This performed the transformation on the front face of the small cube:

   - $[\begin{matrix}0.5 & 0.5 & -4.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}0.5 & 0.5 & 2.6666\end{matrix}]$
   - $[\begin{matrix}0.5 & -0.5 & -4.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}0.5 & -0.5 & 2.6666\end{matrix}]$
   - $[\begin{matrix}-0.5 & -0.5 & -4.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-0.5 & -0.5 & 2.6666\end{matrix}]$
   - $[\begin{matrix}-0.5 & 0.5 & -4.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-0.5 & 0.5 & 2.6666\end{matrix}]$

   and this transformation on the front face of the large cube:

   - $[\begin{matrix}1.0 & 1.5 & -7.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & 1.5 & 6.3332\end{matrix}]$
   - $[\begin{matrix}1.0 & -0.5 & -7.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}1.0 & -0.5 & 6.3332\end{matrix}]$
   - $[\begin{matrix}-1.0 & -0.5 & -7.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & -0.5 & 6.3332\end{matrix}]$
   - $[\begin{matrix}-1.0 & 1.5 & -7.0\end{matrix}]$ $\rightarrow$ $[\begin{matrix}-1.0 & 1.5 & 6.3332\end{matrix}]$

   Here's a render using Blender:

   ![](7a.jpg){width=40%}

   b. \c{(4 points) How would the image change if the camera were moved forward by
   2 units, leaving all of the other parameter settings the same? Determine the
   points, in the image plane, to which each of the cube vertices would be
   projected in this case and sketch the result to scale. Please clearly label
   the coordinates to avoid ambiguity.}

   Here is the updated Blender render:

   ![](7b.jpg){width=40%}

   As you can see, the cubes now take up more of the frame, and in particular
   the red cube has been warped to take up more camera width than the blue.

   c. \c{(4 points) How would the image change if, instead of moving the camera,
   the field of view were reduced by half, to $45^\circ$, leaving all of the
   other parameter settings the same? Determine the points, in the image plane,
   to which each of the cube vertices would be projected and sketch the result
   to scale. Please clearly label the coordinates to avoid ambiguity.}

   Here is the updated Blender render:

   ![](7c.jpg){width=40%}

   Because of the reduced FOV, there is less of the scene shown so the cubes
   take up more of the view. However, there is less of the perspective
   foreshortening effect, so the front cube doesn't get warped into being wider
   or bigger than the back cube.

   d. (2 points)

   - \c{Briefly describe what you notice.}

     The cubes aren't warped except when they change distance from the eye.

   - \c{When looking at two cube faces that are equal sizes in reality (e.g. front
     and back) does one appear smaller than the other when one is more distant
     from the camera than the other?}

     Yes.

   - \c{When looking at two objects that are resting on a common horizontal
     groundplane, does the groundplane appear to be tiled in the image, so that
     the objects that are farther away appear to be resting on a base that is
     higher as their distance from the camera increases?}

     Yes.

   - \c{What changes do you observe in the relative heights, in the image, of the
     smaller and larger cubes as the camera position changes?}

     When the camera position is closer to the cubes, the front cube takes up
     more space overall and so it takes up more height as well. But once the
     camera is far, the big cube has a bigger relative height since their
     heights aren't really warped from each other anymore.

   - \c{Is there a point at which the camera could be so close to the smaller cube
     (but not touching it) that the larger cube would be completely obscured in
     the camera’s image?}

     Yes. You can imagine that if the camera was a microscopically small
     distance from the front cube (and the $near$ value was also small enough to
     accommodate!), then the front cube would take up the entire image.

   - \c{Based on these insights, what can you say about the idea to create an
     illusion of "getting closer" to an object in a photographed scene by
     zooming in on the image and cropping it so that the object looks bigger?}

     It's not entirely accurate, because of perspective warp.

8. \c{Consider the perspective projection-normalization matrix $P$ which maps
   the contents of the viewing frustum into a cube that extends from -1 to 1 in
   $x, y, z$ (called normalized device coordinates).}

   \c{Suppose you want to define a square, symmetric viewing frustum with a near
   clipping plane located 0.5 units in front of the camera, a far clipping plane
   located 20 units from the front of the camera, a $60^\circ$ vertical field of
   view, and a $60^\circ$ horizontal field of view.}

   a. \c{(2 points) What are the entries in $P$?}

   The left / right values are found by using the tangent of the field-of-view
   triangle: $\tan(60^\circ) = \frac{\textrm{right}}{0.5}$, so $\textrm{right} =
   \tan(60^\circ) \times 0.5 = \boxed{\frac{\sqrt{3}}{2}}$. The same goes for the
   vertical, which also yields $\frac{\sqrt{3}}{2}$.

   $$
   \begin{bmatrix}
     \frac{2\times near}{right - left} & 0 & \frac{right + left}{right - left} & 0 \\
     0 & \frac{2\times near}{top - bottom} & \frac{top + bottom}{top - bottom} & 0 \\
     0 & 0 & -\frac{far + near}{far - near} & -\frac{2\times far\times near}{far - near} \\
     0 & 0 & -1 & 0
   \end{bmatrix}
   $$

   $$
   = \begin{bmatrix}
     \frac{2\times 0.5}{\frac{\sqrt{3}}{2} - (-\frac{\sqrt{3}}{2})} & 0 & \frac{\frac{\sqrt{3}}{2} + (-\frac{\sqrt{3}}{2})}{\frac{\sqrt{3}}{2} - (-\frac{\sqrt{3}}{2})} & 0 \\
     0 & \frac{2\times 0.5}{\frac{\sqrt{3}}{2} - (-\frac{\sqrt{3}}{2})} & \frac{\frac{\sqrt{3}}{2} + (-\frac{\sqrt{3}}{2})}{\frac{\sqrt{3}}{2} - (-\frac{\sqrt{3}}{2})} & 0 \\
     0 & 0 & -\frac{20 + 0.5}{20 - 0.5} & -\frac{2\times 20\times 0.5}{20 - 0.5} \\
     0 & 0 & -1 & 0
   \end{bmatrix}
   $$

   $$
   = \boxed{\begin{bmatrix}
     \frac{1}{\sqrt{3}} & 0 & 0 & 0 \\
     0 & \frac{1}{\sqrt{3}} & 0 & 0 \\
     0 & 0 & -\frac{41}{39} & -\frac{40}{39} \\
     0 & 0 & -1 & 0
   \end{bmatrix}}
   $$

   \todo the numbers are wrong lmao

   b. \c{(3 points) How should be matrix $P$ be re-defined if the viewing window
   is re-sized to be twice as tall as it is wide?}

   c. \c{(3 points) What are the new horizontal and vertical fields of view
   after this change has been made?}

## Clipping

9. \c{Consider the triangle whose vertex positions, after the viewport
   transformation, lie in the centers of the pixels: $p_0 = (3, 3), p_1 = (9,
   5), p_2 = (11, 11)$.}

   Starting at $p_0$, the three vectors are:

   - $v_0 = p_1 - p_0 = (9 - 3, 5 - 3) = (6, 2)$
   - $v_1 = p_2 - p_1 = (11 - 9, 11 - 5) = (2, 6)$
   - $v_2 = p_0 - p_2 = (3 - 11, 3 - 11) = (-8, -8)$

   The first edge vector $e$ would be $(6, 2)$, and the edge normal would be
   that rotated by $90^\circ$.

   a. \c{(6 points) Define the edge equations and tests that would be applied,
   during the rasterization process, to each pixel $(x, y)$ within the bounding
   rectangle $3 \le x \le 11, 3 \le y \le 11$ to determine if that pixel is
   inside the triangle or not.}

   b. \c{(3 points) Consider the three pixels $p_4 = (6, 4), p_5 = (7, 7)$, and
   $p_6 = (10, 8)$. Which of these would be considered to lie inside the
   triangle, according to the methods taught in class?}

10. \c{When a model contains many triangles that form a smoothly curving surface
    patch, it can be inefficient to separately represent each triangle in the
    patch independently as a set of three vertices because memory is wasted when
    the same vertex location has to be specified multiple times. A triangle
    strip offers a memory-efficient method for representing connected 'strips'
    of triangles. For example, in the diagram below, the six vertices v0 .. v5
    define four adjacent triangles: (v0, v1, v2), (v2, v1, v3), (v2, v3, v4),
    (v4, v3, v5). [Notice that the vertex order is switched in every other
    triangle to maintain a consistent counter-clockwise orientation.] Ordinarily
    one would need to pass 12 vertex locations to the GPU to represent this
    surface patch (three vertices for each triangle), but when the patch is
    encoded as a triangle strip, only the six vertices need to be sent and the
    geometry they represent will be interpreted using the correspondence pattern
    just described.}

    \c{(5 points) When triangle strips are clipped, however, things
    can get complicated. Consider the short triangle strip shown below in the
    context of a clipping cube.}

    - \c{After the six vertices v0 .. v5 are sent to be clipped, what will the
      vertex list be after clipping process has finished?}

      ![](10a.jpg){width=40%}

    - \c{How can this new result be expressed as a triangle strip? (Try to be as
      efficient as possible)}

      The only way to get this to be represented as a triangle strip is to
      change around the some of the existing lines. Otherwise, the order of the
      vertices prevents the exact same configuration from working.

      See below for a working version (only consider the green lines, ignore the
      red lines):

      ![](10b.jpg){width=40%}

    - \c{How many triangles will be encoded in the clipped triangle strip?}

      Based on the image above, 8 triangles will be used.

## Ray Tracing vs Scan Conversion

11. \c{(8 points) List the essential steps in the scan-conversion (raster
    graphics) rendering pipeline, starting with vertex processing and ending
    with the assignment of a color to a pixel in a displayed image. For each
    step briefly describe, in your own words, what is accomplished and how. You
    do not need to include steps that we did not discuss in class, such as
    tessellation (subdividing an input triangle into multiple subtriangles),
    instancing (creating new geometric primitives from existing input vertices),
    but you should not omit any steps that are essential to the process of
    generating an image of a provided list of triangles.}

    The most essential steps in the scan-conversion process are:

    - First, the input is given as a bunch of geometries. This includes
      triangles and quads. Spheres aren't really part of the primitives, not
      sure why historically but I can imagine it makes operations like
      clipping monstrously more complex.

    - Then, vertices of the geometries are passed to vertex shaders, which are
      custom user-defined scripts that are run in parallel on graphics hardware
      that are applied to each individual vertex. This includes things like
      transforming it through coordinate systems (world coordinates vs. camera
      coordinates vs. normalized coordinates) in order for the rasterizer
      to be able to go through and process all of the geometries quickly.

    - Clipping is done as an optimization to reduce the amount of things outside
      the viewport that needs to be rendered. There are different clipping
      algorithms available, but they all typically break down triangles into
      smaller pieces that are all contained within the viewport. Culling is also
      done to remove faces that aren't visible and thus don't need to be
      rendered.

    - The rasterizer then goes through and gives pixels colors based on which
      geometries contain them. This includes determining which geometries are in
      front of others, usually done using a z-buffer. When a z-buffer is used,
      the color of the closest object is stored and picked. During this process,
      fragment shaders also influence the output. Fragment shaders are also
      custom user-defined programs running on graphics hardware, and can modify
      things like what color is getting output or the order of the depth buffer
      based on various input factors.

12. \c{(6 points) Compare and contrast the process of generating an image of a
    scene using ray tracing versus scan conversion. Include a discussion of
    outcomes that can be achieved using a ray tracing approach but not using a
    scan-conversion approach, or vice versa, and explain the reasons why and why
    not.}

    With ray tracing, the process of generating pixels is very hierarchical.
    The basic ray tracer was very simple, but the moment we even added shadows,
    there were recursive rays that needed to be cast, not to mention the
    jittering. None of those could be parallelized with the main one, because in
    order to even figure out where to start, you need to have already performed
    a lot of the calculations. (For my ray tracer implementation, I already
    parallelized as much as I could using the work-stealing library `rayon`)

    But with scan conversion, the majority of the transformations are just done
    with matrix transformations over the geometries, which can be performed
    completely in parallel with minimal branching (only depth testing is not
    exactly) The rasterization process is also massively parallelizable. This
    makes it faster to do on GPUs which are able to do a lot of independent
    operations.