Real-Time Global Illumination With Radiance Probes | Two Minute Papers #200

Dear fellow scholars, this is two minute papers with this is our 200th episode, so I know you're expecting something great. See how you like this one. One of the most sought after effects in light transport simulations is capturing indirect illumination. This is a beautiful effect where the color of multiple diffuse matte surfaces bleed onto each other. And of course, computing such an effect is as costly as it is beautiful because it requires following the path of millions and millions of light ra. This usually means several hours of waiting time. There have been countless research papers written on how to do this in real time, but the limitations were often much too crippling for practical use. But this time around, you will see soon that these results are just outstanding and we will have a word on limitations at the end of this video. The key contribution of this work is that instead of computing the light transport between all possible point pairs in the scene, it uses radiance probes that measure the nearby illumination and tries to reconstruct the missing information from this sparse set of radiance probes. After that, we place a bunch of receiver points around the scene to places where we would like to know how the indirect illumination looks. There are several things to be taken care of in the implementation of this idea. For instance, in previous works, the hierarchy of these sender and receiver points was typically fixed. In this new work, it is shown that a much sparser set of carefully placed radiance probes is sufficient to create high quality reconstructions. This seemingly small difference also gives rise to a lot of ambiguous cases that the researchers needed to work out how to deal with. For instance, possible occlusions between the probes and receiver points need special care. The entire algorithm is explained in a remarkably intuitive way in the paper. Make sure to have a look at that. And given that we can create images by performing much less computation with this technique, we can perform real time light simulations. As you can see, 3.9 milliseconds is a typical value for computing an entire image, which means that this can be done with over 250 frames per second. That's not only real time, that's several times real time, if you will. Outstanding. And of course, now that we know that this technique is fast, we the next question how accurate is it? As expected, the outputs are always compared to the reference footage, so we can see how accurate the proposed technique is. Clearly there are differences. However, probably many of us would fail to notice that we are not looking at the reference footage, especially if we don't have access to it, which is the case in most applications and note that normally we would have to wait for hours for results like this. Isn't this incredible? There are also tons of more comparisons in the paper. For instance, it is also shown how the density of radiance probes relates to the output quality and where the possible sweet spots are for industry practitioners. It is also tested against many competing solutions. Not only the results, but the number and quality of comparisons is also top tier in this paper. However, like with all research works, no new idea comes without limitations. This method works extremely well for static scenes where not a lot of objects move around. Some movement is still fine as it is shown in the video here, but drastic changes to the structure of the scene, like a large opening door that remains unaccounted for by the probes, will lead to dips in the quality of the reconstruction. I think this is an excellent direction for future research works. If you enjoyed this episode, make sure to subscribe and click the bell icon. We have some more amazing papers coming up. You don't want to miss that. Thanks for watching and for your generous support and I'll see you next time.