In Monte-Carlo based rendering algorithms, the final image is composed of the average value of all samples taken through each pixel. In the "classic" formulation for most algorithms (Path Tracing, Photon Mapping, MTL, etc.) I've always seen that a fixed number or rays is sampled through each pixel. However, I have the impression that clever strategies can be devised to sample a different number of rays per pixel, such that more "troublesome" pixels (which take longer to converge the real average color) receive more samples. My first idea would be to shoot a number of rays proportional to the estimated pixel variance. Is there anyone working on this subject, that can point me to a state of the art survey about such techniques?
I recommend to read "Physically based rendering: from theory to implementation" from Matt Phar. This book contains tons of details about state of art rendering techniques. There is also a section about monte carlo methods, sampling, importance driven sampling and variance. IMHO it make not really sense to reduce the number of rays shooting from the camera as you try to solve the integral over pixels and maybe time.
Hi Falko. Thanks for the answer. I'm aware of existing importance driven sampling techniques, but all approaches I've seen deal with all the samples taken for each pixel separately, and I want to take into consideration the variance of neighboring pixels to decide how to weight the contribution of a given pixel. It seems from your answer that you don't find this useful. I'm totally open to suggestions on this ground, since I'm just grasping the surface here, so please, could you elaborate on why or why not this approach could make sense?
Hi Alejandro. I'm not an expert in this research field! I merely teach the basics and some advanced techniques. I take the most information from the recommended book. The rendering equation implies, that we need to solve an integral for each visible pixel. Due to the fact, that we follow a random walk, we need many rays to reduce the noise in the final image. There exists a couple of techniques dealing with this problem, for instance metropolis light transport. However, core and bones is, that we need to solve an integral. For each surface point we solve the integral to compute the global lighting. For an pixel we solve the integral to reduce (geometric) aliasing, realize camera effects (lens) and support time varying effects like motion blur. I think your idea is based on the fact, that neighboring pixels take similar paths. But in strong mathematical sense, the rendering equation says, that we need to solve the integral for each surface point separately. Maybe i misunderstood your question! Best regards Falko
Hi, not sure if these are quite what you are looking for but "Towards Interactive Global Illumination Effects via Sequential Monte Carlo Adaptation" (see http://www.cs.utah.edu/~vpegorar/research/2008_IRT/) might be helpful. Also the approach by Cline "Table-driven Adaptive Importance Sampling" does something very similar to what you are looking for. Hope it helps!
Thanks! I just download both papers. Seems indeed a lot like what I'm looking for. Will give it a thorough reading as soon as possible, and let you know my impressions.
- Badges
- Science topic
- Similar topics
- Papers