Question: Qualitative results of radiance field modeling

[chaman9000]

Dear authors,

Very interesting work! I am a bit confused with the use of the radiance field. To my understanding, the use of radiance field aims to model the view-dependent effects, such as lighting and reflections. However, I don't see qualitative results related to these types of view-direction-dependent effects.

It seems to me that the view consistent rendering and navigation within the scene can be achieved with a 3D local latent space alone (without the radiance field modeling). It would be great to see the visualization of view-dependent effects that shows the efficacy of the radiance field learning.

[mbautistamartin]

Thanks for your interest! We have now released an official implementation with corresponding training data and pre-trained models so you can check the results by yourself if you want. Indeed one of the advantages of radiance fields is the capacity to model view-dependent effects. I think you are right that you can condition the radiance field in only a 3D point as opposed to a point and a camera direction and see what comes out of it. My guess is that if training data does not contain view-dependent artifacts then it won't make much of a difference.