Volume Rendering, Neural Radiance Fields, and 3D Gaussian Splatting¶

16-825 Learning for 3D: Assignment 3

Author: Benny Zhu (haozhenz) Date: October 2026

Table of Contents¶

  • A. Neural Volume Rendering (60 Points)
    • 1.3 Ray Sampling (5 points)
    • 1.4 Point Sampling (5 points)
    • 1.5 Volume Rendering (20 points)
    • 2. Optimizing a Basic Implicit Volume (10 points)
    • 3. Optimizing a Neural Radiance Field (20 points)
    • 4. NeRF Extras (10 points each)
  • B. 3D Gaussian Splatting (60 Points)
    • 5. 3D Gaussian Rasterization (35 points)
    • 6. Training 3D Gaussian Representations (15 points)
    • 7. Extensions (10 points each)

A. Neural Volume Rendering (60 Points)¶

1. Differentiable Volume Rendering¶

1.3 Ray Sampling (5 points)¶

python volume_rendering_main.py --config-name=box
Grid (xy_grid) Rays
Grid Rays

1.4 Point Sampling (5 points)¶

Sample points

1.5 Volume Rendering (20 points)¶

Color Depth (normalized by max)
Part 1 Depth

2. Optimizing a Basic Implicit Volume (10 points)¶

2.2 Loss and Training (5 points)¶

python volume_rendering_main.py --config-name=train_box
Parameter Value (rounded to 0.01)
Box center (0.25, 0.25, 0.00)
Box side lengths (2.01, 1.50, 1.50)

2.3 Visualization¶

Part 2

3. Optimizing a Neural Radiance Field (20 points)¶

python volume_rendering_main.py --config-name=nerf_lego

Part 3

4. NeRF Extras (10 points each)¶

4.1 View Dependence (10 points)¶

python volume_rendering_main.py --config-name=nerf_materials

Trade-offs: Incorporating view dependence into the NeRF model allows it to capture complex lighting effects, which are essential for reflective surfaces. However, this will also increase the complexity of the model, requiring more parameters and potentially leading to overfitting.


B. 3D Gaussian Splatting (60 Points)¶

5. 3D Gaussian Rasterization (35 points)¶

python unit_test_gaussians.py
python render.py

Part 5 render

6. Training 3D Gaussian Representations (15 points)¶

python train.py
Parameter Learning rate
Opacities 0.01
Scales 0.001
Colours 0.01
Means 0.001

Number of iterations: 500

Metric Value
PSNR 29.355
SSIM 0.937
Training Progress
Final Renders

|

7. Extensions (10 points each)¶

7.1 Rendering Using Spherical Harmonics (10 points)¶

python render.py
Without SH (Q5.5) With SH (Q7.1)

Side-by-side comparisons (same view):

View Without SH With SH
1
2
3

If we render the chair using Spherical Harmonics, the shadow and lighting are better represented as shown in the table. This is because spherical harmonics can capture view dependent effects better, for example, reflections.