Monte Carlo rendering is considered to be de facto standard of photorealistic rendering as it can simulate any light interaction. However, Monte Carlo rendering suffers from noise when there are insufficient samples. The origin of Monte Carlo noise comes from point sampling random parameters which are given as input. In this context random parameter filtering(RPF) identifies Monte Carlo noise and removes them by using joint bilateral filters with consideration of dependency between scene features such as color, world space coordinate, normal, and texture with random parameters which is measured by mutual information, a widely used concept in information theory. RPF shows remarkable result
even when there are samples as small as 8 per pixel without adaptive sampling. However, RPF takes a lot of time because it has to take thousands of neighbor Monte Carlo samples, compute their mutual information, and filter each sample in a pixel one-by-one.
In this paper, we propose Pixel-based random parameter filtering. Instead of running random parameter filering algorithm per sample, we modified it to run per pixel. Calculation of mutual information, which has to be done sample-by-sample because of the necessity to analyze effects of random parameters on each sample, can be significantly reduced by taking superresolution approach. Our results show orders of magnitude speed acceleration with little loss in quality.