**Assignment 3**
Student name: Gaurav Parmar
(#) 1. Differentiable Volume Rendering
(##) 1.3 Ray Sampling
The xy grid (left) and rays (right) are shown below
(##) 1.4 Point Sampling
The point samples are visualzed below.
(##) 1.5 Volume Rendering
The final resulting gif (left) and depth (right) are visualized below.
(#) 2. Optimizing a basic implicit volume
(##) 2.1 Random ray sampling
The corresponding function is implemented in the accompanying code.
(##) 2.2 Loss and training
The final center of the box is found to be:
(0.25, 0.25, -0.00)
The side lengths found are:
(2.00, 1.50, 1.50)
(##) 2.3 Visualization
I obtain a similar resulting cube as the reference in the writeup.
The gif of the obtained results is shown below.
(#) 3. Optimizing a Neural Radiance Field (NeRF)
A NeRF trained on the bulldozer is shown below.
The model is optimized for 30 epochs.
I use hormonic embedding for the input and a 4 layer MLP with LeakyReLU non linearity.
(#) 4. NeRF Extras
(##) 4.1 View Dependence
The low resolution nerf model trained with view dependence is shown below.
The model is optimized only for 10 epochs because of extended runtime.
Note that the view angle should be concatenated only with color mlp, and not for density mlp.
(##) 4.3 High Resolution Imagery
The high resolution nerf model is shown below.
The model is optimized for just 10 epochs because of extended runtime.
(#) late days
I regretably used 1 late days on this assigment.
