Enabling multispectral rendering about redner HOT 2 OPEN

Thanks for the explanation. I am considering parameter estimation problems using plenoptic and multispectral observations. Ability to compute the derivatives of the observations w.r.t key scene parameters can help us build estimators (and also quantify fundamental limits) for such problems.

I am happy to provide more information and specifics about these kinds of problems if you'd like. Please do let me know if these type of problems interests you (or your group).

from redner.

Currently we assume we sample three points in the spectrum (usually RGB but it really can be anything). Changing this requires one to replace all the "Vector3" in the C++ code to a spectrum type like what is done in pbrt and mitsuba. There are also some tricks that sample different part of the spectrum for each light path sample (e.g. the "Hero" wavelength sampling method), which is not too hard to add. Dispersion and other wavelength-dependent effects require more changes.

A common hack (also supported in Mitsuba) is to convert the spectral data into RGB space for material parameters and emission before rendering. This can be easily done in PyTorch.

May I ask what are the applications you are looking at? I might not have time to do this in short term but I can put this into my TODO list.

from redner.