16-825 Assignment 1: Rendering Basics with PyTorch3D

1. Practicing with Cameras

1.1 360-degree Renders

Cow turntable

1.2 Re-creating the Dolly Zoom

Dolly zoom


2. Practicing with Meshes

2.1 Constructing a Tetrahedron

Tetrahedron

4 vertices and 4 triangular faces.

2.2 Constructing a Cube

Cube

8 vertices and 12 triangular faces.


3. Re-texturing a mesh

Retextured cow

I choose color1 = (0, 0, 1) (blue) for the front of the cow, the vertex with the smallest z, and color2 = (1, 0, 0) (red) for the back


4. Camera Transformations

First Transform

t1

R_relative = Rz(-90), T_relative = (0, 0, 0).

R_relative rolls the camera 90 degrees but T_relative does nothing. The camera stays where it is and the cow rotates.

Second Transform

t3

R_relative = I, T_relative = (0, 0, 2).

To make the cow smaller, I defined R_relative as doing nothing where as T_relative put the camera 2 units back.

Third Transform

t4

R_relative = I, T_relative = (0.5, -0.5, 0).

The cow is facing slightly left but the pose remains the same. So R_relative should be doing nothing and tweak T_relative.

Forth Transform

t2

R_relative = Ry(90), T_relative = (-3, 0, 3).

Ry(90) rotate the camera by 90 degree. But without a corresponding change to the translation, the cow completely falls outside the view. So we need to move camera back using T_relative.


5. Rendering Generic 3D Representations

5.1 Rendering Point Clouds from RGB-D Images

RGB-D point clouds

5.2 Parametric Functions

Torus point cloud

Cube point cloud

5.3 Implicit Surfaces

Torus mesh

Ellipsoid mesh

The new object is an ellipsoid, (x/a)^2 + (y/b)^2 + (z/c)^2 - 1 = 0, with semi-axes a, b, c = 0.4, 0.8, 1.2.

Mesh vs point cloud: trade-offs

Rendering quality. Comparing two donuts side by side, the mesh is having a better quality as the surface is generally smooth. It's easy to see holes for the point cloud one and we can sometime even look through it.

Rendering speed. Comparing the render speed: the mesh is faster of the two, by 2.5x. Measured on the same torus at 256x256 on CPU: 1.42 s per frame for the mesh against 3.60 s for the point cloud.

Ease of use. We can easily get point clouds from sensors, just like what we did for the plant. But in terms of rendering as mesh vs. as point cloud, I think it is not so much different.

Memory. For this torus the point cloud is 938 KB and the mesh 754 KB. The implicit function is a few bytes of code, but discretizing it costs O(n^3) memory to make a 3D grids: the 64^3 grid alone is 1024 KB, larger than either thing it produces.


(Extra Credit) 6. Do Something Fun

Dragon

The XYZ RGB Dragon (249,882 triangles) from the Stanford 3D Scanning Repository, via Alec Jacobson's common-3d-test-models.