1
$\begingroup$

I'm currently trying some things with transmittance estimators in PBRT and I have implemented the following paper: Unbiased Ray-Marching Transmittance Estimator. This method assumes that two points $x$ and $y$ are given.

In PBRT and tracking methods, a distance can be sampled analytically due to null-collisions and a homogenized medium (i.e. delta tracking). The pdf $p(t)$ for this distance $t$, is known in a homogenized medium thanks to the null-scattering path integral formulated by Miller et al. This enabled the use of MIS in heterogeneous media, since the path pdf is analytically available.

In my integration scheme however, I do not know the pdf for the distance $t$. I am aware of the following sampling scheme where the pdf of a distance $t$ is given by:

$$ p(t) = \frac{dF(t)}{dt} = \frac{d}{dt}(1 - e^{-\tau(t)}) = \frac{d\tau(t)}{dt}T(t) = \mu_t(t)T(t)$$

Here $F(t) = 1 - T(t)$ is the CDF of sampled distances, $\tau(t)$ is the optical thickness a distance $t$ away from $x$ in the direction of $y$ and $T(t)$ is the transmittance. Finally $\mu_t(t)$ is the medium extinction at a distance $t$ from $x$ (see Monte Carlo methods for Volumetric Light Transport).

Can I simply compute the transmittance for a given distance $t$ and use this (unbiased) transmittance value in the computation of the path pdf? I'm hesitant to do so since this value for transmittance is not the real value but an approximation.

$\endgroup$

1 Answer 1

1
$\begingroup$

To remain unbiased, the PDF $p(x)$ in $\sum f(x) / p(x)$ should be the PDF of the actual sample distribution. That is, the only PDF you can use depends on how you get your samples.

So, for example (I didn't read the paper, I am just sharing one of the possible way I think feasible), you use ray marching and subdivide the ray into several bins before calculating the transmittance and local $\mu_t$, and within the sampled bin, you resort to uniform sampling for simplicity. Since ray marching just approximates the continuous PDF with a discrete one, you might have the following sampling PDF: $$ p(t_{\text{within-bin}} | k)p(k) = \frac{1}{\text{bin width}}\mu_t(t_k)T(t_k)/Z $$ Here we use $k$ to represent the k-th sampled bin. Therefore, if unbiasedness holds:

  • Your sampling method should be sampling a bin first according to normalized $\mu_t(t)T(t)$, which is $p(k)$ and within that bin, uniformly sampling the final distance value. Then your PDF should be exactly the same form (in the equation above) since this is defined by the sampling method.
  • Make sure your PDF is normalized. This is easy for inverse-transform sampling but may be hard for other method (that requires analytical integration to get $Z$, the normalization factor).

So even if you say this is approximated, it would work in a unbiased way if the rules are followed. The accuracy of approximation will only affect the convergence speed (zero-variance, if $f(x) = c\times p(x), c$ is a constant).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.