Question about "vis_mask"

[wen-yuan-zhang]

Thanks for your excellent work!
I noticed that you used visible mask to modify the indicator function in Eq.(9) in the paper, which is written as
$$\Psi (t)=\prod (1 - h(t) * m(t))$$
However, it seems that the implementation code (Line 410 in ''udf_renderer_blending.py'') is written as
$$\Psi (t)=\prod (1 - h(t) + flip\_saturation*m(t))$$,
and the variable $flip\_saturation$ increases from 0 to 1 along with the training progress. Would you please explain why using ''add'' instead of ''multiply'' to deal with the indicator function? And what is the role of $flip\_saturation$ here?

Thanks again!

[xxlong0]

Thank you for pointing it out.
Ideally, we can multiply with the vis_mask to get the unbiased indicator function.
In the implementation, we adopt a softer practice. Instead of applying multiply, we add the vis_mask and then clip to (0,1):
(1. - alpha_occ + saturation*vis_mask).clip(0, 1) .
The implementation has some advantages:
The saturation parameter can soften the masking operation, allowing gradients for the "masking" points when saturation is smaller than 1.
This is because, in the early stage of the optimization, the vis_mask might be not very accurate, so we can use the saturation parameter to control the masking operation. Only at the end of optimization, we mask out the points before the surfaces.