UC Berkeley

Computer graphics rendering is undergoing a renaissance, with physically-based rendering methods based on accurate Monte-Carlo image synthesis replacing ad-hoc techniques in a variety of applications including movie production. In interactive applications like product visualization or video games, physically-based lighting effects are increasingly popular. However, producing photo-realistic images at interactive speeds still remains a challenge. In Monte-Carlo rendering, a pixel’s color is computed by sampling and integrating over a high dimensional space. This includes effects like (1) motion blur, due to objects moving during the time the camera shutter is open; (2) defocus blur, due to camera lens optics; (3) area and environment map lighting, which is direct illumination coming from many directions; (4) global illumination, due to light reflected from one surface to another. The color is sampled through ray- or path-tracing. With insufficient rays, the image looks noisy because the integrand has high variance, and 1000s of rays are needed (per pixel) for a pleasing image. Previous work has showed a Fourier analysis for some of these effects, deriving a compact double-wedge spectrum, and a sheared filter that aligns with the slope of the spectrum. This filter can remove noise from a very sparsely sampled Monte-Carlo image, but is very slow. In this thesis, we will extend the Fourier analysis for more general cases, and propose a less compact axis-aligned filter, that aligns with the frequency axes. The resulting spatial bandwidths are then used for image-space filtering, that is orders of magnitude faster than sheared filtering. The packing of the Fourier spectra also provides adaptive sampling rates that minimize noise in conjunction with the adaptive filter. These algorithms improve speed relative to converged ground-truth by about 30x-60x, and we are able to demonstrate interactive speed with a GPU ray-tracer. We also demonstrate an application of our method to mixed reality with a Kinect camera.