Assignment 1: Rendering Basics with PyTorch3D
16-825 Learning for 3D Vision | Fall 2026 | Jinyao Zhou (jinyaozh)
1.1 360-degree Renders (5 points)
1.2 Re-creating the Dolly Zoom (10 points)
2.1 Constructing a Tetrahedron (5 points)
The mesh has 4 vertices and 4 triangle faces.
2.2 Constructing a Cube (5 points)
The mesh has 8 vertices and 12 triangle faces (2 triangles per cube side).
3. Re-texturing a mesh (10 points)
color1 = [1, 0, 0] (red)
color2 = [0, 0, 1] (blue)
4. Camera Transformations (10 points)
R_relative = [[0, 1, 0], [-1, 0, 0], [0, 0, 1]]
T_relative = [0, 0, 0]
R_relative = I
T_relative = [0, 0, 2]
R_relative = I
T_relative = [1/3, -1/3, 0]
R_relative = [[0, 0, 1], [0, 1, 0], [-1, 0, 0]]
T_relative = [-3, 0, 3]
5.1 Rendering Point Clouds from RGB-D Images (10 points)
Left to right: the point cloud from the first image, from the second image, and the union of the
two.
5.2 Parametric Functions (10 + 5 points)
Torus
Heart
5.3 Implicit Surfaces (15 + 5 points)
Torus
Mesh vs. point cloud.
- Rendering speed. Point clouds are cheaper: splatting N points needs no
connectivity and no shading model. A mesh pays for a rasterisation and a shader, and getting the mesh
here also costs a marching-cubes pass over a dense 128³ grid, which dominates the runtime.
- Rendering quality. A mesh surface is continuous and opaque, however, the details like sharp
edges might be not well represented, unless setting high voxel resolution. A point cloud is usually full of empty space
unless heavily oversampled.
- Memory usage. A point cloud stores 3 floats per sample and no connectivity; a mesh adds a
face list, and an implicit one needs the whole voxel grid before extraction. But matching mesh-level
quality takes far more points, might need some experiments to find the best balance case by case.
- Ease of use. Point clouds can be read directly from sensors. Meshes contains more geometry information
like normals, surface area, watertightness, collision, etc. which are not well defined on a point cloud.
Heart
(Extra Credit) 7. Sampling Points on Meshes (10 points)
Left to right: the original cow mesh, then point clouds of 10, 100, 1000 and 10000 sampled
points.