Assignment 1: Rendering Basics with PyTorch3D

16-825 · Learning for 3D Vision

1. Practicing with Cameras

1.1 360° Renders

360 degree turntable render of the cow mesh

1.2 Re-creating the Dolly Zoom

Dolly zoom effect on the cow mesh

2. Practicing with Meshes

2.1 Constructing a Tetrahedron

360 degree render of a tetrahedron
Vertices: 4 Faces: 4

2.2 Constructing a Cube

360 degree render of a cube
Vertices: 8 Faces: 12

3. Re-texturing a Mesh

Cow mesh retextured with a front-to-back color gradient

color1 = [0, 0, 1] and color2 = [1, 0, 0]

4. Camera Transformations

Transform 1

My reproduction of transform 1

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

This R_relative is a −90° rotation about the camera's z-axis, with no translation change

Transform 2

My reproduction of transform 2

R_relative = identity, T_relative = [0,0,3]

With no rotation, adding 3 units to the z-translation pushes the camera 3 additional units back along its viewing axis, roughly doubling the camera-to-object distance and making the cow appear smaller and farther away.

Transform 3

My reproduction of transform 3

R_relative = identity, T_relative = [0.5,-0.5,0]

With no rotation, the x-component of T_relative shifts the camera 0.5 units sideways, which moves the cow horizontally in the frame, while the negative y-component shifts the camera 0.5 units downward, which makes the cow appear lower in the frame.

Transform 4

My reproduction of transform 4

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

R_relative is a +90° rotation about the camera's y-axis, swapping the x and z axes so the camera looks at the cow from the side. T_relative then repositions the camera along these new axes to bring it closer and keep the cow centered, producing a close side-profile view of the cow.

5. Rendering Generic 3D Representations

5.1 Rendering Point Clouds from RGB-D Images

Point cloud from first RGB-D image
Image 1 point cloud
Point cloud from second RGB-D image
Image 2 point cloud
Union of both point clouds
Union of both point clouds

5.2 Parametric Functions

Torus point cloud (10 pts)

360 degree render of a torus point cloud
Torus point cloud (R = 2, r = 1), 200×200 samples

Extra object of choice (+5 pts)

360 degree render of a superellipsoid point cloud
Superellipsoid point cloud (Barr parametrization, e1 = e2 = 0.4)

5.3 Implicit Surfaces

Torus mesh (15 pts)

360 degree render of a torus mesh from an implicit function
Torus mesh via marching cubes (R = 2, r = 0.6)

Mesh vs. point cloud tradeoffs:
Rendering Speed: Point clouds are generally simpler to render because each point can be rendered independently. Meshes require rasterizing triangles and handling their connectivity, which can be more computationally expensive.
Rendering quality: Meshes usually produce higher-quality and more continuous-looking surfaces because triangles explicitly define the surface between vertices. They also support smooth shading, textures, and lighting naturally. Point clouds can look sparse or have visible gaps, when the number of points is small.
Ease of use: Point clouds are easier to construct because we only need a set of 3D coordinates. Meshes require both vertices and faces, so we need to keep track of how vertices are connected to form surfaces.
Memory usage: Point clouds only need to store the coordinates and features of each point, whereas meshes need to store both vertex data and face connectivity. Thus, meshes generally require more memory.
Representation and editing: Meshes provide explicit surface and connectivity information, which makes them better for representing continuous surfaces. Point clouds are more flexible for representing raw 3D measurements, such as LiDAR or RGB-D data, where surface connectivity may not be available.

Extra object of choice (+5 pts)

360 degree render of a superellipsoid mesh from an implicit function
Superellipsoid via marching cubes, ((x/1.5)²)^(1/4) + ((y/1.2)²)^(1/4) + ((z/1.0)²)^(1/4) − 1 = 0

6. Do Something Fun extra credit

360 degree render of a trefoil knot mesh
Trefoil knot, rendered as an implicit surface

7. Sampling Points on Meshes extra credit

Cow mesh next to a 10-point sample
10 points
Cow mesh next to a 100-point sample
100 points
Cow mesh next to a 1,000-point sample
1,000 points
Cow mesh next to a 10,000-point sample
10,000 points