
These are my visualizations of the xy pixel grid and the ray directions.

These are my sample points along the rays.

This is my spiral rendering of the box on the left and the normalized depth map from the third camera on the right.

This is my spiral rendering of the optimized box. After sampling and training, the values I got were:
- center: (0.25, 0.25, -0.00)
- side lengths: (2.01, 1.50, 1.50)

This is my NeRF. This version depends only on position. The way it works is that a 3d position is passed through a positional encoding and then through a 6 layer MLP which outputs density and color.

This is the modified NeRF to add view dependance. I did this by fedding the viewing direction into the color part of the network but not the part that predicts density, to ensure the shape stays the same from all angles.
The trade off is that this view dependance can hurt generalization. If the color depends too much on the view, the network will just memorize how the training image looks instead of actually learning the shape.

This is my render of the pre-trained chair Guassians.


These are my training progress gifs and final renders after training. These were my hyperparams and results:
- learning rates: opacities = 0.05, scales = 0.005, colours = 0.02, means = 0.001
- number of iterations: 1000
- mean PSNR: 29.38
- mean SSIM: 0.938

This is my render using spherical harmonics (the section 5 render is without it). Below are some side-by-side comparisons.

This is frame 3. With SH (right), the seat and the backrest seem to have more shading and the gold pattern is a little brighter. The chair also looks more full.

This is frame 14. With SH (right), the seat has shading that changes with the viewing angle, while without it, the seat is evenly shaded.