Assignment 1: Rendering Basics with PyTorch3D¶
Name: Chenzhe Zhang
Course: 16-825 Learning for 3D Vision
1.1 360-degree Renders¶
1.2 Re-creating the Dolly Zoom¶
3. Re-texturing a Mesh¶
I chose color1 = [0, 0, 1] (blue) for front and color2 = [1, 0, 0] (red)for back.
4. Camera Transformations¶
transform1¶
The relative rotation rotates the cow's projection clockwise by 90 degrees, so its head points to the right. The relative translation is zero, leaving the camera distance unchanged.
$$ R_{\mathrm{relative}} = \begin{bmatrix} 0 & 1 & 0 \\ -1 & 0 & 0 \\ 0 & 0 & 1 \end{bmatrix} $$
$$ T_{\mathrm{relative}} = [0,0,0] $$
transform2¶
The relative rotation is the identity matrix, preserving the cow's orientation. The relative translation increases the camera-space depth from 3 to 6, making the cow appear smaller.
$$ R_{\mathrm{relative}} = \begin{bmatrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix} $$
$$ T_{\mathrm{relative}} = [0,0,3] $$
$$ T = [0,0,6] $$
transform3¶
The relative rotation is the identity matrix. In PyTorch3D's camera coordinates, positive x moves the projection to the left, and negative y moves it downward. The depth remains unchanged.
$$ R_{\mathrm{relative}} = \begin{bmatrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix} $$
$$ T_{\mathrm{relative}} = [0.5,-0.5,0] $$
$$ T = [0.5,-0.5,3] $$
transform4¶
The relative rotation rotates about the y-axis to show the cow from the side, with its head on the left. The relative translation compensates for the rotated initial translation and keeps the cow centered.
$$ R_{\mathrm{relative}} = \begin{bmatrix} 0 & 0 & 1 \\ 0 & 1 & 0 \\ -1 & 0 & 0 \end{bmatrix} $$
$$ T_{\mathrm{relative}} = [-3,0,3] $$
The final translation is:
$$ T = R_{\mathrm{relative}}T_0 + T_{\mathrm{relative}} = [3,0,0] + [-3,0,3] = [0,0,3] $$
5.1 Rendering Point Clouds from RGB-D Images¶
From left to right: the point cloud from image 1, the point cloud from image 2, and their union.
5.3 Implicit Surfaces¶
Torus¶
Mesh vs. Point Cloud¶
Meshes produce smooth, connected surfaces, while sparse point clouds may show gaps. Point clouds are easier to create because they do not need faces. Rendering speed and memory usage depend on the number of points or triangles and the rendering settings.
Additional Object: Ellipsoid¶
6. Do Something Fun¶
I designed three color schemes for a sofa, assigning different vertex colors to the cushions, frame, and legs. I rendered a synchronized 360-degree comparison of the three designs.
Model: Sofa by berkayc on Free3D.
7. Sampling Points on Meshes¶