Assignment 1: Rendering Basics with PyTorch3D

Name: Tashvik Dhamija

1.1 360-degree Renders

Cow 360
Cow mesh, 360° turntable

1.2 Re-creating the Dolly Zoom

Dolly zoom
Dolly zoom effect

2.1 Constructing a Tetrahedron

Tetrahedron 360
Tetrahedron, 360° turntable

Vertices: 4  |  Faces: 4

2.2 Constructing a Cube

Cube 360
Cube, 360° turntable

Vertices: 8  |  Faces: 12

3. Re-texturing the Cow

Retextured cow
Front-to-back color gradient

color1 = [0, 0, 1] (blue) is assigned to the front of the cow (smallest z), and color2 = [1, 0, 0] (red) to the back (largest z), linearly interpolated per vertex by z-coordinate.

4. Camera Transformations

Base view
Base view (R_0, T_0)

Transform 1

Transform 1
R_relative rotates 90° about the z-axis

R_relative = [[0,1,0],[-1,0,0],[0,0,1]], T_relative = [0,0,0].
A pure roll: the camera rotates 90° about its own viewing axis (z), so the cow appears rotated within the image plane. No translation needed.

Transform 2

Transform 2
R_relative rotates 90° about the y-axis, with T_relative repositioning the camera

R_relative = [[0,0,1],[0,1,0],[-1,0,0]], T_relative = [-3,0,3].
The camera yaws 90° about the y-axis to look from the side, and T_relative moves the camera to a new position beside the cow so it stays centered and in front of the lens. Pure rotation alone would put the cow out of the camera's view frustum.

Transform 3

Transform 3
T_relative translates the camera away from the cow

R_relative = identity, T_relative = [0,0,2].
Pure translation along the camera's viewing axis. The camera moves farther from the cow, so it appears smaller (zoomed out). No rotation.

Transform 4

Transform 4
T_relative shifts the camera in x and y

R_relative = identity, T_relative = [0.5,-0.5,0].
Pure lateral translation. The camera shifts sideways and vertically, changing the framing without any rotation or change in distance.

5.1 Rendering Point Clouds from RGB-D Images

Plant image 1
Point cloud from image 1
Plant image 2
Point cloud from image 2
Plant union
Union of both point clouds
Plant side by side
All three, side-by-side

5.2 Parametric Functions

Torus point cloud
Torus (required)
Helix point cloud
Helix (bonus, +5)

5.3 Implicit Surfaces

Torus mesh
Torus mesh, via marching cubes (required)
Figure-8 mesh
Figure-8 / dumbbell mesh (bonus, +5)

Mesh vs. Point Cloud: Tradeoffs

BasisMesh (implicit surface)Point Cloud (parametric)
Rendering speedSlower to build. Requires voxelizing a full 3D grid and running marching cubes, cost scales with voxel_size³.Fast. Points come directly from evaluating a parametric equation on a 2D grid, no voxelization step.
Rendering qualityContinuous, smoothly shaded surface since triangles interpolate normals across their area. Looks solid even at moderate resolution.No true surface. Gaps show through at low sample density and there is only per-point color, no continuous shading.
Ease of useMore setup: define an implicit function, choose a bounding box and voxel resolution, run marching cubes, then build the mesh.Simpler: define the parametric equations and sample directly, no extra extraction step.
Memory usageFace count depends on marching cubes resolution and can grow quickly (voxel_size³), even for simple shapes.Memory scales directly and predictably with the number of points sampled.

Extra Credit 6: Do Something Fun

Baymax
Baymax, built from 6 parametric ellipsoids (head, torso, 2 arms, 2 legs) combined into one scene

Extra Credit 7: Sampling Points on Meshes

Original cow mesh
Original mesh
10 points
N = 10
100 points
N = 100
1000 points
N = 1,000
10000 points
N = 10,000

Points are sampled via area-weighted face selection (torch.multinomial over per-face triangle area) followed by a uniform random barycentric coordinate on the chosen face, so larger triangles receive proportionally more samples.