Q1.3¶
| vis_grid | vis_rays |
|---|---|
 |
 |
Q1.4¶
Q1.5¶
| gif | depth |
|---|---|
 |
 |
Q2.3¶
Box center: (0.250, 0.250, 0.000)
Box side lengths: (2.005, 1.503, 1.503)| my output | TA's output |
|---|---|
 |
 |
Q3¶
Following is the output of the default setting nerf_lego.
Q4.1¶
In this part I show the different results with vanilla network and the network that encode ray directions.
| Vanilla Network | View point dependence |
|---|---|
 |
|
In order to show more details. I also generate the output with higher resolution, here I used the training setup of Q4.3.
| Vanilla Network | View point dependence |
|---|---|
 |
 |
The overfitting effect is not obvious in both case. I think this may because the lego case is not sufficient enough to represent more view changes.
Q4.3¶
In this part I am curious about the difference of sampling different number of points in one ray. In this section, I compared the output from 64 points per ray and 128 points per ray. With 128 points sampled per ray, the output model have a better result in some details, i.e., the wheel in the back.
Due ot the GPU memory limitation, I didn't generate the result with more points. The result of the experiment is listed below.
| points per ray | 64 | 128 |
|---|---|---|
| output |  |
|