1.3. Ray sampling
1.4. Point sampling
1.5. Volume rendering
Box center: (0.250, 0.250, -0.0002) Box side lengths: (2.004, 1.503, 1.503)
The network is 4 layer MLP with hidden dim 512. On training for 50 epochs the MSE loss decreases to 0.005042 and the visualization looks like:
The shared network is a 4 layer MLP which outputs feature and density. feature and viewdir are then used in another 4 layer MLP to predict color. On training for 50 epochs the MSE loss decreases to 0.005042 and the visualization looks like:
Adding view direction dependency on the color values makes it model reflection and other specularities but results in less generalizability. It can lead to overfitting to the given views. We observe floating artifacts at the bottom of the rendered views from novel directions.
For highres training I increased the n_harmonic_functions_xyz to 12.
Directly training on the architecture similar to lowres resulted in slightly low quality rendering. The visualization looks like this:
When using a deeper MLPWithInputSkips network with 10 layers, 256 hidden dimension and skip connection at 4th layer to train the model. Resulting model has improved rendering quality.
While keeping the same architecture as above and increasing the n_pts_per_ray doesnt have a significant effect on the NeRF rendering in this case. But the training time increases linearly with it and thus takes longer.
