Question about light direction estimation about differentiable-shadow-mapping HOT 5 CLOSED

Hi,

do you mind sharing your modifications? Which example did you modify?

For the light pose estimation, I simply set up a direction vector with requires_grad = True, and in each iteration use it to create a DirectionalLight instance that is passed to the SimpleRenderer.render. Also, don't forget to scale the direction vector to length 1 in each iteration.

from differentiable-shadow-mapping.

Thank you for your help. I know how to fix the issue.
The point is to use it to create a DirectionalLight in each iteration.

If I create the DirectionalLight as a global variation, the issue occurs.

from differentiable-shadow-mapping.

Hm, interesting. Still, the issue should not occur. Did you setup the optimizer to optimize DirectionalLight.direction?

from differentiable-shadow-mapping.

In fact, I create a direction with nn.Paramters Class and setup a corresponding optimizer like
direction = torch.tensor([[0.57, 0.57, 0.57]], dtype=torch.float32, device=device, requires_grad=True) and
optimizer = torch.optim.Adam([direction], lr=0.3).

In each iteration, if I create light as light = DirectionalLight(direction=F.normalize(direction)) in the loop, it is OK, but if I create light before the training begins, the issue occurs.

from differentiable-shadow-mapping.

Anyway, the issue is solved, thanks again for your interesting work and help.

from differentiable-shadow-mapping.