Assignment 1: Rendering Basics with PyTorch3D

1. Practicing with Cameras

1.1 360-degree Renders

cow

1.2 Re-creating the Dolly Zoom

dolly zoom

2. Practicing with Meshes

2.1 Constructing a Tetrahedron

tetrahedron

4 vertices, 4 triangular faces

2.2 Constructing a Cube

cube

8 vertices, 12 triangular faces

3. Re-texturing a Mesh

Front (color1) is yellow [1, 1, 0], back (color2) is blue [0, 0, 1]

retextured cow

4. Camera Transformations

Transform 1

R_relative rolls the camera 90 degrees and T_relative is zero

t1

Transform 2

R_relative turns the camera 90 degrees around the vertical axis and T_relative keeps the cow centered

t2

Transform 3

R_relative has not been changed and T_relative moves the camera farther back, making the cow smaller

t3

Transform 4

R_relative has not been changed and T_relative shifts the cow left and down in the image

t4

5. Rendering Generic 3D Representations

5.1 Point Clouds from RGB-D data

point cloud from first image, second image, and union

5.2 Parametric Functions

Torus

rotating torus point cloud

Cone

rotating cone point cloud

5.3 Implicit Surfaces

Torus

rotating torus mesh

Rounded cube

rotating round cube mesh

A mesh has a continuous surface with lighting and smooth edges but the point cloud needs enough samples and a good point radius if we want to avoid visible gaps. The mesh takes more time to extract with marching cubes and stores both faces + vertices. A point cloud can be simpler and cheaper to create if its sparser, but if it's pretty densely sampled the points can cost more to render because their projected disks overlap. It's usually better to use meshes if the surface will be viewed many times and point clouds are usually better when the samples are already available.

7. Sampling Points on Meshes

increasing # of sampled points and then the original:

point clouds and mesh