Is it possible to make sdf calculation differentiable?

[yangyichu]

Say I have a point p=(xyz), and a distance d = sdf(p), I want to do something like d.backward() to get p.grad? I was able to do it with TORCH_SDF library(a modified version of Kaolin), but it seems I cannot do it with pytorch_volumetric?

[liboyan233]

I think you may implement this with the sdf_grad in:
sdf_val, sdf_grad = sdf(pts)
However, instead of apply sdf_grad to sdf_val, it should be applied to pts directly. This may because the sdf function is rather implemented in an implicit way, which means there's no analytical differential relationship between sdf_value and pts.

[patricknaughton01]

Just wanted to bump this because I agree with @yangyichu that this would be a great feature. It seems like it's currently possible to manage the sdf_grad back propagation manually but it would be much more unified and convenient if this were handled automatically.

[LemonPi]

@yangyichu @liboyan233 @patricknaughton01 Can you guys try out #10 (autograd branch) to see if it addresses your needs and if it breaks anything? I've added tests and it seems to be working.

[patricknaughton01]

Just tested it and it works for my use case (keeping an end effector a certain distance from a mesh). No errors on my end.

[LemonPi]

Nice, will add some documentation and then merge if no one else has issues