16-889 Assignment 1: Rendering Basics with PyTorch3D

0. Main script Usage

main.py script arguments

usage: main.py [-h] [-q {q1.1,q1.2,q2.1,q2.2,q3,q4,q5.1.1,q5.1.2,q5.1.3,q5.2,q5.3,q6,q7}]
               [--obj_path OBJ_PATH] [-i IMAGE_SIZE] [-n NUM_SAMPLES] [-o OUTPUT_PATH] [-f NUM_FRAMES]
               [-d DURATION]

optional arguments:
  -h, --help            show this help message and exit
  -q {q1.1,q1.2,q2.1,q2.2,q3,q4,q5.1.1,q5.1.2,q5.1.3,q5.2,q5.3,q6,q7}
  --obj_path OBJ_PATH
  -i IMAGE_SIZE, --image_size IMAGE_SIZE
  -n NUM_SAMPLES, --num_samples NUM_SAMPLES
  -o OUTPUT_PATH, --output_path OUTPUT_PATH
  -f NUM_FRAMES, --num_frames NUM_FRAMES
  -d DURATION, --duration DURATION

1. Practicing with Cameras

1.1. 360-degree Renders (5 points)

Run command :

python main.py -q q1.1 --obj_path data/cow.obj -i 256 -n 100 -o output/q11.gif  -f 30 -d 2

Q1.1

1.2 Re-creating the Dolly Zoom (10 points)

Run command :

python main.py -q q1.2 --obj_path data/cow_on_plane.obj -i 256 -n 100 -o output/q12.gif  -f 30 -d 2

Q1.2

2. Practicing with Meshes

2.1 Constructing a Tetrahedron (5 points)

Run command :

python main.py -q q2.1 -i 256 -n 100 -o output/q21.gif  -f 30 -d 2

Cow render

Number of Vertices = 4
Number of faces = 4

2.2 Constructing a Cube (5 points)

Run command :

python main.py -q q2.2 -i 256 -n 100 -o output/q22.gif  -f 30 -d 2

Cow render

Number of Vertices = 8
Number of faces = 12

3. Re-texturing a mesh (10 points)

Run command :

python main.py -q q3 --obj_path data/cow.obj -i 256 -n 100 -o output/q3.gif  -f 30 -d 2

Cow render

I used color1 = [0, 0, 1] and color2 = [1, 0, 0]

4. Camera Transformations (20 points)

Run command :

python main.py -q q4 --obj_path data/cow.obj -i 256 -o output/q4

R_relative	T_relative	Description
[[0, 1, 0], [-1, 0, 0], [0, 0, 1]]	[0, 0, 0]	Rotate the camera clockwise 90° along z-axis
[[1, 0, 0], [0, 1, 0], [0, 0, 1]]	[0, 0, 2]	Camera is moved 2 units in -z direction to get the zoomed out effect.
[[1, 0, 0], [0, 1, 0], [0, 0, 1]]	[0.5, -0.5, 0]	Move the camera 0.5 units in -x axis and 0.5 units in y axis
[[0, 0, 1], [0, 1, 0], [-1, 0, 0]]	[-3, 0, 3]	Rotate the camera anti-clockwise 90° in y axis and then move 3 units in the new x-axis and -3 units in the new z-axis

5. Rendering Generic 3D Representations

5.1 Rendering Point Clouds from RGB-D Images (10 points)

Run command :

python main.py -q q5.1.1 -i 256 -o output/q51_a.gif 
python main.py -q q5.1.2 -i 256 -o output/q51_b.gif  
python main.py -q q5.1.3 -i 256 -o output/q51_c.gif

Plant1	Plant2	Both

5.2 Parametric Functions (10 points)

Run command :

python main.py -q q5.2 -n 100 -i 256 -o output/q52_100.gif 
python main.py -q q5.2 -n 1000 -i 256 -o output/q52_1000.gif

100 points	1000 points

5.3 Implicit Surfaces (15 points)

Run command :

python main.py -q q5.3 -i 256 -o output/q53.gif

Cow render

Comparision between Mesh and Point cloud

To have a continuous 3d surface representation, point cloud requires a lot of points to fill any kind of gaps. While, a mesh always has a continuous representation. So visually rendering quality of mesh is better.
In the above example, we needed (1000^2 = 1e6) points to represent the torus using parametric form and we only needed (64^3 ~ 2.6e5) for implicit representation to construct the mesh. We observe that the time taken to render point cloud is more compared to render the mesh.
To get the continuous representation, we need to store less number of (vertex + face) data compared to point cloud. So memory requirement is less for mesh.
In case of ease of use, I think point cloud is easier to handle. If we want to deform the object, since every point is independent in point cloud, its easier to experiment with. In case of mesh, due to strong dependence with the neighboring points that form face, we need to manipulate more data.

6. Do Something Fun (10 points)

Run command :

python main.py -q q6 -i 256 -o output/q6.gif

I have downloaded snowflake 3D model from free3D. I duplicated this mesh 25 times at random positions. To render the image, I moved my camera diagonally to get this snowfall effect :) I noticed pixelating artifacts on rendering an image of 256x256 resolution. So I increased the resolution to 1024x1024.

Cow render

Here is an another experiment where I moved camera towards the snowflakes

Cow render

(Extra Credit) 7. Sampling Points on Meshes (10 points)

Run command :

python main.py -q q7 -i 256 -n 10 -o output/q7_10.gif 
python main.py -q q7 -i 256 -n 100 -o output/q7_100.gif 
python main.py -q q7 -i 256 -n 1000 -o output/q7_1000.gif 
python main.py -q q7 -i 256 -n 10000 -o output/q7_10000.gif

Cow mesh	Points : 10	Points : 100	Points : 1000	Points : 10000