How to get the pixel color if a training ray it traverses multiple submodules?

[Chen Wang]

This work is excellent. But I have a question, in the paper, you said each submodule can be trained separately, but what if a training ray corresponds to multiple submodules, do you merge the sample points in different modules?

[hturki]

at train time, we'll include the ray in the train set of all cells it traverses (since we don't know the 3D structure of the scene a priori). at inference time, we'll query the submodule of the cell each sampled point is within before compositing into a single pixel value