16-825 Assignment 1: Rendering Basics with PyTorch3D

1.1 360-degree Renders

1.2 Dolly Zoom

2.1 Tetrahedron

4 vertices, 4 faces.

2.2 Cube

8 vertices, 12 faces.

3 Re-texturing a Mesh

color1 = orange [1.0, 0.5, 0.0] at the front (smallest z), color2 = lime [0.75, 1.0, 0.0] at the back (largest z).

4 Camera Transformations

  1. R_relative rotates the camera 90° about the z axis. T_relative = 0.
  2. R_relative rotates the camera 90° about the y axis; T_relative = [-3, 0, 3] moves it to the side of the cow at the same distance.
  3. R_relative = identity; T_relative = [0, 0, 2] moves the camera 2 units further back along z.
  4. R_relative = identity; T_relative = [0.5, -0.5, 0] shifts the camera right and up.

5.1 Point Clouds from RGB-D Images

Left: image 1. Middle: image 2. Right: union.

5.2 Parametric Functions

Torus, and a Möbius strip.

5.3 Implicit Surfaces

Torus mesh, and a heart mesh.

For the torus, rendering the point cloud with 200×200 samples (40k points total) took about 3.3 s per frame. After converting it to a mesh with marching cubes, the result had about 7k vertices and 14k faces and rendered in around 1 s per frame, so the mesh was roughly 3× faster. Visually, the mesh also looks nicer because it is a continuous surface, while the point cloud render contain holes. If we want to have less holes from the point cloud render we would have to sample more points which further slows the rendering.

The main advantage of the point cloud was that it was easier to generate. For rendering meshes, it requires evaluation of a function and then run marching cubes. I also had to make sure the grid covered enough space to cover the complete surfae

In terms of memory, there was not a huge difference at this scale. The point cloud stores 40k 3D positions, while the mesh stores about 7k vertices plus 14k faces. The mesh does require extra connectivity information and the temporary voxel grid during construction.

6 Do Something Fun

10000 points sampled on the cow, each given a random velocity and pulled down by gravity.

7 Sampling Points on Meshes

Left to right: 10, 100, 1000, 10000 sampled points, and the original mesh.