Depth differentiability

[amefabris]

Hi everyone,
I am trying to understand whether the output of the mesh resterizer, more specifically fragments.zbuf[0, ..., 0] is differentiable or not. From the discussion in issue #35 it seem that it IS differentiable.

Nevertheless, my doubt comes from the fact is that if we look at fragments.zbuf[0, ..., 0] of a random object as below

and I zoom in to try to visualize the differentiability, I only see sharp boundaries.

The same does NOT apply when I visualize the silhouette extracted from the SoftSilhouetteShader, in this case the silhouette looks and is differentiable.

Thank you in advance for any insights!

[bottler]

It is differentiable to some changes in the input but not others (e.g. object boundaries moving), and this can be combined with soft blending/ shading to get a more differentiable result.

[amefabris]

Thank you very much for your prompt response! Could you elaborate more on what you mean by "object boundaries moving"? And how can I leverage soft blending/shading to get a more differentiable results on the depth map?
Currently what I'm doing is (following what was posted in issue #35 ):
I created a custom class named Renderer which takes as input mesh, camera, blend_params, rast_settings, and has in its constructor among other things:

self.rasterizer = MeshRasterizer(cameras=self.cameras, raster_settings=self.rast_settings)
self.shader = SoftSilhouetteShader(blend_params=self.blend_params)

Then in the forward function I do:

fragments = self.rasterizer(self.meshes, R=R, T=T, cameras=self.cameras)
image = self.shader(fragments, self.meshes, R=R, T=T)
return [ [fragments.zbuf, image] ]

But fragments.zbuf is not affected by either shading or blending...

[amefabris]

I found out that in the latest release there is a SoftDepthShader, I'll try to use that.
Posting here in case anyone is in the same situation as me.