Speaker: Adam Celarek (ICGA)
For non correlated rendering algorithms like path tracing or bidirectional path tracing a simple and effective convergence, or error, metric is the variance. As soon as the variance becomes small, we know that the error is small. This is not true for algorithms, where samples of the integral are correlated to each other. Metropolis Light Transport (MLT) is such an algorithm. Once a region with low sampling probability and high contribution is found, the algorithm keeps sampling from the same region and neglects other parts of the image, resulting in a non stratified overall sampling.
We want to find a measure to describe this behaviour and we want to formulate a description of a fingerprint, so that we can compare different rendering algorithms efficiently.