mitsuba3
mitsuba3 copied to clipboard
about the parameter to_world of camera or projector emmits
Hi, There is a confused question I would like to ask. I check the camera and projetor plugin, I fount the to_world params can not be differential , So would you like to give me some advice about how can I optimize the camera pose or projector pose in mitsuba Here I find other people do projector pose estimation using params "Projector.position" and "Projector.rotation" But I cant find these params in my program. Thanks.
the code some researcher does: params = traverse(scene) params.keep(["Projector.position", "Projector.rotation"])
Hi @ljijjiang1234
Currently the differentiable integrator do not properly handle discontinuities (e.g. using reparameterization) at the sensor level. For this reason we decided to keep the to_world
parameter non-differentiable for now to avoid users working with incorrect gradients.
I should be able to take the time to look into this in the next couple weeks. I will keep you posted once this is fixed.
Thank you very much about your work , And I Will Try to use the your base work to compile and redefine the emitter source code in order to make it differentiate like someone else already done.
The last few commits on master
(ef9f559e0989fd01b43acce90892ba9e0dea255b .. ea513f73b65b8776afb75fdc8d40db4b1140345e) have made sensors parameters differentiable.
Note: Emitters still don't have a differentiable to_world
parameter I believe.
Hi could someone guide me how exactly I can make my to_world params differentiable in case of an orthographic sensor?
Mitsuba 3 cannot differentiate a orthographic sensor correctly.
Let me close this issue. The original question has been answered.