Assignment 3

1. Differentiable Volume Rendering

1.3. Ray sampling

Visualization of xy_grid and rays:

1.4. Point sampling

Visualization of point samples from camera:

1.5. Volume rendering

Visualization of rendered box and depth image:

2. Optimizing a basic implicit volume

2.3. Visualization

Center of Box: (0.25, 0.25, -0.00)

Side Length: (2.00, 1.50, 1.50)

Reconstructed box using MSE Loss:

3. Optimizing a Neural Radiance Field (NeRF)

Reconstructed low-res lego bulldozer:

4. NeRF Extras

4.1. View Dependence

Reconstructed low-res lego bulldozer with view-dependence:

Increased view-dependence into training the model seems to add more texture in certain areas that would have been smoothed out if not using view-dependence.
I did not notice that view-dependece takes longer to converge and train. Increased view-dependence also overfits the model to that specific scene which decreases generalization on novel scenes.

4.3. High Resolution Imagery

Reconstructed high-res lego bulldozer with view-dependence:

I trained the high resolution model for 300 epochs which took much longer than the low-resolution. I increased the number of layers to 8 and the append channel to 4.
I had to decrease the chunck size by a factor of 2 to 8192 for GPU contraints. I also increased the number of points per ray to 180.
Overall I notice much more detail in the lego blocks at the base of the bulldozer.