Data-driven multiple importance sampling for monte carlo rendering

Abstract

Monte Carlo (MC) rendering is the most common method in Computer Graphics to generate photo-realistic images. To enhance it efficiently, a variety of methods related to sampling has been developed, one of the simple and efficient way is Multiple Importance Sampling (MIS). This technique is a powerful technique for combining several sampling strategies. However, choosing an optimal weight in combining several sampling strategies remains a challenge. To address this problem, we propose a data-driven weight computation for reducing the variance of MIS. The point of our method is to utilize the scene information. We use precomputation for utilizing scene information, and it allows for computing an optimal weight depending on a scene. Specifically, optimal weight varies on each portion of the scene. Our method applies an optimal weight to an image locally. We observed meaningful results over prior methods in different scenes.