Assignment 3

A. Neural Volume Rendering

1. Differentiable Volume Rendering

1.3 Ray Sampling


XY grid visualization

Ray bundle visualization

1.4 Point Sampling


Sampled Points visualization

1.5 Volume Rendering


Volume rendering visualization

Depth Visualization

2. Optimizing a basic implicit volume

2.2 Loss and training

Box side lengths: (2.005, 1.503, 1.503)
Box center: (0.250, 0.250, -0.000)

2.3 Visualization


Optimized Volume

3. Optimizing a NeRF


NeRF Visualization

4. NeRF Extras

4.1 View Dependence


NeRF with view dependence

B. 3D Gaussian Splatting

5. 3D Gaussian Rasterization

5.5 Perform Splatting


Splatting Visualization

6. Training 3D Gaussian Representations

6.2 Perform Forward Pass and Compute Loss

Learning rates used

pre_act_opacities: 0.004
pre_act_scales: 0.0009
colours: 0.004
means: 0.0004

Iterations: 1000
Mean PSNR: 27.02
Mean SSIM: 0.915


Training Progress

Splatting Visualization

7. Extensions

7.1 Rendering Using Spherical Harmonics


Splatting Without Spherical Harmonics

With Spherical Harmonics

Without Spherical Harmonics

With Spherical Harmonics
The image rendered with spherical harmonics has a lot more depth to it. In the image rendered without spherical harmonics, the cushions are pretty flatly colored compared to the other image. View-dependent appearance is improved by allowing the color of the gaussian to vary with the viewing direction.

Without Spherical Harmonics

With Spherical Harmonics
Even in this image, its visible how much the view dependence adds to the quality of the color of the chair rendered. The back cushion is colored completely differently, with different shadows. The accents on the chair are also different.

Without Spherical Harmonics

With Spherical Harmonics