Assignment 1: Rendering Basics with PyTorch3D

Name: Sumanth Yegurla · Andrew ID: syegurla

0. Setup

Rendering the provided cow mesh with a point light.

Rendered cow mesh

1.1 360-Degree Renders

A 360-degree turntable gif of the cow mesh, generated using look_at_view_transform over 36 evenly spaced azimuth values.

Cow mesh 360-degree turntable

1.2 Re-creating the Dolly Zoom

The dolly zoom effect, moving the camera back while increasing the field of view so the cow stays the same size in frame.

Dolly zoom effect

2.1 Constructing a Tetrahedron

Vertices: 4 · Faces: 4

Tetrahedron 360-degree turntable

2.2 Constructing a Cube

Vertices: 8 · Faces: 12 (2 triangles per cube face)

Cube 360-degree turntable

3. Re-texturing a Mesh

The cow mesh re-textured with a front-to-back color gradient based on each vertex's z-coordinate. color1 = (0, 0, 1) (blue) is assigned to the front of the cow (z_min), and color2 = (1, 0, 0) (red) is assigned to the back (z_max), linearly interpolated in between.

Re-textured cow mesh

4. Camera Transformations

Reference render with the original camera extrinsics R_0, T_0:

Original camera view
Transform 1 result
Transform 1 — R_relative is a 90° rotation about the camera's Z (viewing) axis, rolling the cow onto its side. T_relative = [0, 0, 0] (no translation).
Transform 2 result
Transform 2 — R_relative is a 90° rotation about the camera's Y (up) axis, yawing to a side profile of the cow. T_relative shifts the camera to move the cow further away and into the lower-left of the frame.
Transform 3 result
Transform 3 — R_relative is the identity (no rotation). T_relative is a pure translation along Z, pushing the camera further from the cow.
Transform 4 result
Transform 4 — R_relative is the identity. T_relative is a pure translation along X, shifting the cow toward the right of the frame.

5.1 Rendering Point Clouds from RGB-D Images

Point clouds unprojected from the plant RGB-D data: image 1, image 2, and their union, rendered side-by-side as a 360-degree turntable.

Plant point clouds, side by side turntable

5.2 Parametric Functions: Torus

A torus point cloud sampled from its parametric surface function, rendered as a 360-degree turntable with the hole clearly visible.

Torus point cloud 360-degree turntable

Bonus: New Parametric Object — Mobius Strip

A Mobius strip point cloud sampled from its parametric function of u (angle around the loop) and v (position across the width), rendered as a 360-degree turntable.

Mobius strip point cloud 360-degree turntable

5.3 Implicit Surfaces: Torus Mesh

A torus mesh recovered via marching cubes on the implicit function F(x, y, z) = (sqrt(x² + z²) - R)² + y² - r², rendered as a 360-degree turntable with the hole clearly visible.

Torus mesh 360-degree turntable

Mesh vs. Point Cloud Tradeoffs

Bonus: New Implicit Object — Metaball Snowman

A "snowman" mesh built from 5 metaball fields (3 stacked body segments plus 2 arms), each a compact-support field f_i(p) = (1 - (d/r_i)²)³ for d < r_i, summed and thresholded to blend the segments together, recovered via marching cubes and rendered as a 360-degree turntable.

Metaball snowman mesh 360-degree turntable

Extra Credit: Do Something Fun

A textured MQ-9 Reaper drone model, rendered as a 360-degree turntable.

Textured MQ-9 drone 360-degree turntable

Source

Extra Credit: Sampling Points on Meshes

Points sampled uniformly from the surface of the cow mesh via stratified sampling: a face is chosen with probability proportional to its area, then a uniform random barycentric coordinate on that face gives the sampled point. Shown side-by-side, left to right: the original mesh, then point clouds of 10, 100, 1000, and 10000 points.

Cow mesh and point-sampled clouds at increasing density, side by side turntable