Physically Accurate Noise Free Real-time Rendering

Abstract

This is the master thesis project performed by Mauro van de Vlasakker under supervision of Dr. Robby T. Tan at Utrecht University. Physically based rendering is a widely studied topic in the field of computer graphics. The goal is to find the average color of each pixel by solving the rendering equation. The rendering equation describes light transport mathematically. Path tracing is a way to solve the rendering equation and is done by shooting rays from the camera through each pixel on the screen. Many rays are needed to get a good estimate for the final pixel color. If the number of rays per pixel is low, this will show up as noise in the image. This project is focused on post processing the output image where the goal is to make the post process fast enough to enable real-time (24 fps) path traced scenes. C++ and OpenCL are used to implement a path tracer which is used as a platform to implement the filter. Random parameter filtering is implemented to directly compare our filters quality with theirs. The main problems we solve are: real-time performance, filtering reflections/ refractions and filtering complex Monte Carlo effects like soft shadows and depth of field. Using the variance in scene information we can detect noise. By splitting direct and indirect illumination we can apply separate filtering on each to achieve better results. Skylights and sky-boxes are taken into account since they usually have no normals and are infinitely far away which can cause the filter to over blur parts of the scene. We evaluate the speed of the filter at different resolutions with different test scenes. The results show that the speed of the filter is fast enough to achieve real-time performance and scales linearly with resolution. Then, the filter is carefully evaluated to determine the optimal filtering parameters for all test scenes. With the optimal parameters we compare our filter against Random Parameter Filtering in terms of quality. We show that our filter is able to filter reflection/refractions, depth of field, soft shadows and sky-boxes/skylights in real-time. For future work there are some interesting extensions that can be made such as adaptive sampling and detecting when the filtering process can stop.