Since the depth tensor is generated in the projection kernel and UT does not implement a backward pass, the depth gradient does not directly contribute to the gaussian means.
Ideally, the depth gradient is directly applied in the eval3D backward kernel, but if I am not mistaken, that does not happen.
Losses on depth will not have the same affect using ut=eval3D=False as with ut=eval3d=True.
Since the depth tensor is generated in the projection kernel and UT does not implement a backward pass, the depth gradient does not directly contribute to the gaussian means.
Ideally, the depth gradient is directly applied in the eval3D backward kernel, but if I am not mistaken, that does not happen.
Losses on depth will not have the same affect using ut=eval3D=False as with ut=eval3d=True.