16889-Assigntment3

Jinkun Cao (jinkunc@andrew.cmu.edu)

1. Differentiable Volume Rendering

1.3 Ray Sampling

The output grid visualization and ray visualization are as below.

1.4 Point Sampling

The output visualization of render points is shown as below

1.5 Volume Rendering

The output rendered volume is shown as below as well as the followed depth.

2. Optimizing a basic implicit volume

The code implementations are available in the submitted files.

2.2 Loss and training

The box center is (0.250, 0.250, -0.000), the box side lengths are (2.004, 1.501, 1.503)

2.3 Visualization

The visualization of the optimized volume is as below

3. Optimizing a Neural Radiance Field (NeRF)

The visualization of the output of the implemented NeRF without view dependence is shown as below

4. NeRF Extras

4.1 View Dependence

The output by NeRF with view dependence and from the same hyperparameter setting is shown as below

To make the view-depdencen implementation, I simply concatenate the embedding of direction features for those sampled points and the features from their positions for the color generation layers. Because the input dimension of the linear layers is increased, it brings some more computation overhead, though minor for this case. The generalization quality is a little boosted, but not significant for this sample case.