Assignment 3
Anirudh Chakravarthy (achakrav)
Late days: 0
Question 1
Question 1.3
Grid/ray visualizations:
Question 1.4
Point samples:
Question 1.5
Color/depth visualizations:
Question 2
Question 2.2
Box center: (0.2502. 0.2506, -0.0005)
Box side lengths: (2.0051, 1.5036, 1.5034)
Question 2.3
Volume visualization:
Question 3
I used positional encoding but no view dependence.
Question 4
Question 4.1
I followed the view dependence implementation from the NeRF paper.
On adding view dependence, I observed two key differences:
- The patterns at the base of the bulldozer seem more clear during rendering. In the paper, they share this observation, since view dependence allows the network to handle specularities better.
- Without view dependence, we observe some artifacts in the rendering. Specifically, when the bulldozer (in Q3) is turned away from us, we can momentarily observe a reflection below the base of the bulldozer. This is not present on adding view dependence.
Based on these observations, I believe that adding view dependence improves generalization ability to unseen views. This allows the network to render realistic colours consistent with viewing directions.
Question 4.3
For training a high-resolution NeRF, I experimented with 3 settings
Baseline
In this setting, I trained the same network (as in the previous questions) for high-resolution images.
This resulted in a blank render with a black background. This led to the conclusion that the network capacity was not sufficient to learn.
Deeper network
In this setting, I followed the exact network from the NeRF paper (depth, skip connections, hidden dimensions, etc). I used a batch size of 256 and chunk size of 8192.
As seen in the GIF above, the network is able to render fine-grained details (the dotted pattern) on the base on the bulldozer. This was absent in the previous results.