16-825 Assignment 1: Rendering Basics with PyTorch3D

1. Practicing with Cameras

1.1. 360-degree Renders

360-degree cow mesh render

1.2. Dolly Zoom

Dolly zoom effect

2. Practicing with Meshes

2.1. Tetrahedron

360-degree tetrahedron render

4 vertices and 4 triangular faces.

2.2. Cube

360-degree cube render

8 vertices and 12 triangular faces.

3. Re-texturing a Mesh

color1 = [0, 0.5, 1] (blue) at the front and color2 = [1, 0.6, 0] (orange) at the back. The contrasting colors are interpolated along the z-axis.

360-degree cow render with blue to orange texture

4. Camera Transformations

R₀ = I and T₀ = [0, 0, 3], where I is the identity matrix. The new camera uses R = R_relative @ R₀ and T = R_relative @ T₀ + T_relative.

Target 1: Image roll

Image roll
R_relative = [[0, 1, 0], [-1, 0, 0], [0, 0, 1]]
T_relative = [0, 0, 0]

R_relative rolls the image 90° clockwise. T_relative leaves the camera translation unchanged.

Target 3: Move farther away

Move farther away
R_relative = I
T_relative = [0, 0, 3]

R_relative preserves orientation. T_relative increases the depth from 3 to 6, making the cow appear smaller.

Target 4: Shift left and down

Shift left and down
R_relative = I
T_relative = [0.5, -0.5, 0]

R_relative preserves orientation. T_relative shifts the cow left and down: positive camera x points left, and positive camera y points up.

Target 2: Side view

Side view
R_relative = [[0, 0, 1], [0, 1, 0], [-1, 0, 0]]
T_relative = [-3, 0, 3]

R_relative rotates the view 90° around the vertical axis. T_relative compensates for the rotated initial translation, keeping the final T = [0, 0, 3] and the cow centered.

5. Rendering Generic 3D Representations

5.1. Point Clouds from RGB-D Images

First image
First image
Second image
Second image
Union
Union

5.2. Parametric Functions

Torus point cloud
Torus point cloud
Additional object: Möbius strip
Additional object: Möbius strip

5.3. Implicit Surfaces

Torus mesh
Torus mesh
Additional object: blended spheres
Additional object: blended spheres

Meshes give continuous surfaces and support lighting, while sparse point clouds can leave holes. Point clouds are easy to build directly from RGB-D images or parametric samples; meshes require face connectivity or surface extraction. Rendering speed depends on triangle or point count and screen coverage; dense point clouds can be expensive. Meshes store vertices, colors, and face indices, while point clouds store positions and colors but may need many samples for comparable coverage. Implicit mesh extraction also needs a temporary voxel grid.

7. Sampling Points on Meshes (Extra Credit)

10, 100, 1,000, and 10,000 sampled points beside the original cow mesh