cgtuebingen / robust-vmm-guiding Goto Github PK

I am wondering if there is a tool to visualize the guiding field similar to PPG

Adding the Teaser image of the paper to the readme.

At the moment the (max) num of SPP per training iteration is not configurable via the GUI and set to 4spp.
In addition, the behavior of the SPP per iteration is unintuitive (4, 8, 16). The latter is due to being able to
still compare exponential and incremental training.

To make it possible to also use incremental learning with 1SPP and being able to show that our algorithm is still robust (even when setting the num of VMM initial components to 1 :)), this option should be made available again.
Also, the behavior should be changed to always using the number of set samples (getting rid of the exponential behavior).
To be able to still compare inc vs exp training a boolean flag can be added (we do not need to expose this to the GUI since this experiment is only for internal use) to switch the behavior of the max num spp per iteration parameter.

I also think it would make be useful to mention the ability to do incremental training with 1SPP and even 1 component in the readme (with a small example comparing it to the settings in the paper).
This "really" proves the robustness of our method and is a big selling point.

Thanks for your great work.
As described above, is this method useful for light map baking? In light map baking, the distance between revels is large due to the resolution of light map and the distribution of radiance isn't narrow enough (smooth), so is this method still helpful for baking?

Thanks for releasing the code. I am unable to access scene files using the link provided. Can you kindly provide the updated link?

Hello, I enjoy reading your awesome paper on path guiding.
However, there is one detail that I don't understand well. The following sentence appears in the Hemispherical Guiding with Spherical Distributions subsection: 'The probability of such samples can be significantly reduced by computing the product with the surface-normal oriented cosine-lobe just before sampling.' I don't understand exactly what it means to compute the cosine before performing sampling. Don't we have to sample the direction vector to calculate the cosine? I'm curious how the normal vector and cosine were used to truncate only the upper hemisphere part from the von-Mises Fisher mixture distribution and sample it unbiasedly.