16-825: Learning for 3D Vision — Assignment 1¶

Shengxi Wu


1.1. 360-degree Renders¶

1.2 Re-creating the Dolly Zoom¶

2.1 Constructing a Tetrahedron¶

4 Vertices & 4 Faces

2.2 Constructing a Cube¶

8 Vertices & 12 Faces

3. Re-texturing a mesh¶

I selected:

color1=[1.0, 0.41, 0.71] (pink)

color2=[0.58, 0.25, 0.85] (purple)

4. Camera Transformations¶

Pytorch performs camera transform with: x_cam = Rᵀ x_world + T

R_relative rotates the camera orientation, which determines the direction it is faced towards, with it initially facing the +z direction. For instance, with transform 1 it is a 90 degree rotation around the world's z-axis to rotate the view of the cow.

T_relative shifts the position of the camera after rotation, thus making it a translation in the camera's own coordinate frame rather than the world's. The camera ends up displaced by the negative of that value.

Transform 1:¶

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

Transform 2:¶

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

Transform 3:¶

R_relative = I, T_relative = [0,0,2]

Transform 4:¶

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

5.1 Rendering Point Clouds from RGB-D Images¶

(Left: The point cloud corresponding to the first image1, Middle: The point cloud corresponding to the second image, Right: The point cloud formed by the union of the first 2 point clouds)

5.2 Parametric Functions¶

Torus

The figure 8 immersion of Klein Bottle, equation found in Wikipedia

Source: https://en.wikipedia.org/wiki/Klein_bottle

5.3 Implicit Surfaces¶

Torus

A surface of Genus2, equation found in Wikipedia

Source: https://en.wikipedia.org/wiki/Implicit_surface

Tradeoffs between rendering as a mesh vs a point cloud:¶

Rendering speed: Rendering speed of point cloud increase with number of points, whereas rendering speed of mesh increase with rendering method, number of samples per pixel, which is likely the same regardless how complex the mesh is, disregarding loading time.

Rendering quality: Generally, mesh has better rendering quality, even with less triangles in a mesh, it still has less holes than a denser pointcloud due to presence of faces, making the continuous surface and holes easier to distinguish.

Ease of use: Pointcloud since there is no connectivity information, so it's easier to up/down sample, modify (delete points), which is more more complicated for mesh since it may have to go through re-shape and change of connectivity relationship for and modification.

Memory usage: Mesh is generally larger for the same amount of points (vertices) as its point cloud equivalence, but its quality would be much worse.

Generation: To generate a point cloud from RGBD data is much simpler than mesh, as point cloud would only require projection of RGB information to its corresponding depth, whereas to generate a mesh an additional poisson reconstruction would be necessary. But for artists or architects, it is easier to generate mesh from scratch when drawing with CAD.

Generally, for purpose that is more practical, like 3D printing or light-tracing based rendering, mesh is necessary; but for simpler task like alignment or perspective match point cloud is much simpler to use and generate.

6. Do Something Fun¶

Arc shoot - Basically a dolly zoom and an orbit run at the same time.Used by filmmakers to build tension in a shot.

7. Sampling Points on Meshes¶

Sample 10 points

Sample 100 points

Sample 1000 points

Sample 10000 points