Study On Efficient Sampling Methods For Realistic Rendering In Complex Scenes

Photorealistic rendering is one of the most important research topics in computer graphics. It is widely used in design and entertainment industries, and requires accurate simulation of light transport process to render a realistic image of the virtual scene. Most photorealistic rendering methods are based on numerically solving an integral equation. Samples are randomly generated in the light path space, then weighted sum to yield an approximate solution of the integral equation. The light path space is high-dimensional and very complex, and the computation cost of these methods is also high. We aim at improving the performance of photorealistic rendering of complex scenes by proposing a series of efficient sampling techniques.One idea is to increase sample utilization for improving rendering performance. Along this direction, we present two photon mapping based realistic rendering methods. The first one aims at global illumination rendering, and uses irradiance regression to improve sample utilization in final gathering. The non-parametric regression model combines advantages of photon mapping and final gathering, and predicts irradiance of final gathering by direct irradiance estimation from photon mapping. Using this method, we are able to achieve similar quality of final gathering at the cost of direct photon mapping. To address failed predictions, we propose to cull outliers by a fast pruning algorithm, and shade these outlier samples by final gathering. Under similar visual quality, this method achieves about 3 times overall speedup comparing to final gathering. The second method aims at efficient rendering of inhomogeneous translucent materials. We improve utilization of photon samples by adjusting probability of photon distribution. This method is based on volumetric photon mapping, and efficiently increases the number of photons inside view volume by adjusting photon distribution probability, and therefore improves translucent rendering quality. To solve the photon power differences caused by probability adjustment, we propose to split photons in the dimensions of distance and angle, respectively. High power photons are split into photons with lower power. Under same rendering cost, our results are superior than that of volumetric path tracing and volumetric photon mapping.Another idea is reducing the computation cost of a single sample, and this paper combines new parallel computing mode with massive parallel computing on GPU to improve rendering performance. We propose an efficient scalable programmable motion effects algorithm. Motion effects significantly visualize object motion hint using motion depiction styles, and are an extension of photographic blur. We present a parallel rendering system on GPU for programmable motion effects, and parallelize computation in both temporal and image dimensions. This new parallel mode significantly improves data reuse among threads, and better balances GPU memory usage and parallelism of programmable motion effects rendering. To resolve visibility conflicts between trace segments, we introduce a new visibility algorithm based on interval split. For the convenience of programmable motion effects design, we precompute trace segments for each pixel, to further reduce rendering cost.