academic discussion about unboundednerfpytorch

[JonasW-byte]

For example, we split the San Francisco dataset as training dataset and test dataset, and train the training dataset as one block, can we get better result on test dataset?

[sjtuytc]

This is a good question. In fact, the generalization of NeRF is a hard topic from my point of view. In the real world, two different places may have different physical properties (reflexity, density, etc). However, the physical laws (including the rendering law, Newton's laws, etc) keep constant when going from one place to another. We cannot guarantee that we can generalize the NeRF from one place to another. However, when testing NeRF performances, we can still sample from the local block so that we can generalize from training views to unseen views.

[sjtuytc]

Closed because of no further comment.

[JonasW-byte]

Yes, exactly. Except reflection, density, a lots of factors have influence on the generalization of NeRF. Movable object(vehicles, pedestrians, etc.) is also one of factors? because the collected sensors take photos at one place with different time in my dataset, which causes same place with different feature(color). sometime with movable object, sometimes not. If we remove the object like waymo, that is a expensive engineer.
Btw, 2D rendering head maybe easier than depth output head? Because sometime we want to get reconstruct 3D scene from depth info. It's hard to keep consistent on same object and get better result on whole image.

[sjtuytc]

Yes, you are absolutely correct.