Point-wise Adaptive Filtering for Fast Monte Carlo Noise Reduction

Abstract

Monte Carlo based photorealistic image synthesis has proven to be one of the most flexible and powerful rendering techniques, but is plagued with undesirable artifacts known as Monte Carlo noise. We present an adaptive filtering method designed for Monte Carlo rendering systems that counteracts noise while respecting sharp features. The filter operates as a post-process on a noisy image augmented with three screen-space geometric attribute buffers, and by using a point-wise adaptive (varying window size) filtering kernel, this method is able to reinforce the preservation of important scene reflected edges, in less time. Comparative results demonstrate the simplicity and efficiency of our method, which makes it a feasible and robust solution for smoothing noisy images generated by Monte Carlo rendering techniques. CUDA implementation also makes the algorithm potentially practical for interactive Monte Carlo rendering in the near future.