Assignment 3

1. Differentiable Volume Rendering

You can run the code for part 1 with:

python main.py --config-name=box

Output:

Grid	Rays

1.4. Point sampling (10 points)

Run:

python main.py --config-name=box

Output:

1.5. Volume rendering (30 points)

Run:

python main.py --config-name=box

Output will be stored as images/part_1.gif

3D 360 View	Depth

2. Optimizing a basic implicit volume

Run:

python main.py --config-name=train_box

Output:

3. Optimizing a Neural Radiance Field (NeRF) (30 points)

Run:

python main.py --config-name=nerf_lego

Output:

4. NeRF Extras (Choose at least one! More than one is extra credit)

4.1 View Dependence (10 pts)

First, modify the config file by setting implicit.view_dependence = 1

Then, run:

python main.py --config-name=nerf_lego

Output:

Without View Dependence	With View Dependence

As shown above, view dependence changed the color for some parts of the legos. Specifically, the image seems to be a bit brighter yellow compared to with view dependence. In general, adding view dependence will result in our output quality being better, however, this comes at a cost of overfitting to the specific scene and not being able to generalize to other similar scenes.

4.3 High Resolution Imagery (10 pts)

Then, run:

python main.py --config-name=nerf_lego_highres

Output:

Low Res	High Res

Point Samples Per Ray	Output
32
128
256

Visual quality is better as the number of points is increased up to a certain point. However, as the number of points per ray increases, the time to train the model also increases.