Course: 16-889 Learning for 3D Vision

Name: Soyong Shin

Due by: Mar. 17 (Thu)

I used 5 late day for this assignment

Contents:

1. Differentiable Volume Rendering

1.3. Ray sampling (10 points)

Code implementation at ray_utils.py

1.4. Point sampling (10 points)

Code implementation at sampler.py

1.5. Volume rendering (30 points)

Code implementation at renderer.py

2. Optimizing a basic implicit volume

2.1. Random ray sampling (5 points)

Code implementation at ray_utils.py

Randomization was implemented using torch.randperm

2.2. Loss and training (5 points)

Code implementation at main.py

$Loss = \frac{1}{N} \sum_{i=1}^{N} ||F^{gt}[i] - F^{pred}[i]||_2$

2.3. Visualization

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

Code implementation at implicit.py and main.py

• Network training is highly susceptible to the initialization of layers. I use xavier_uniform initialization which helped the stable training.

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

4.1. View Dependence (10 pts)

To use view dependent variable, viewpoint (ray bundle’s direction) is embedded using harmonic embedding. Then it is fused with positional embedding. The code is implemented at implicit.py.

4.2. High Resolution Imagery (10 pts)