Light Transport Matrix Based Photorealistic Rendering

Recently,the photorealistic rendering has become more and more popular in many applications.As an important technique for photorealistic rendering,the global illumina-tion technique is a fundamental research in a very long time.It aims to vividly stimulate the position,the shape,the material,the participating media,and most importantly,the shading and lighting,of the real world.For all applications require high-quality photore-alistic images,such as designing,movie creation,game development,virtual reality,etc,global illumination is an important pillar.People proposed various solution to try to effi-ciently solve this time-consuming problem.A promising one is the many-light rendering method that draws increasing attention in recent years.It represents the multiple-bounces light transport with virtual lights to efficiently reuse the light sub paths.It is not only effi-cient,but also flexible for making a trade-off between performance and quality.However,there are still some challenges,including how to efficiently gather virtual point lights,how to efficiently render large-scale out-of-core scene,and how to efficiently calculate the contributions of camera sub paths.This thesis mainly focus on the formulation,sampling and reconstructing light transport matrix to efficiently render high-quality photorealistic images.It includes a novel sparse sampling and recovery approach that adapts matrix separation method to accelerate the general many-light rendering,an adaptive sampling and matrix comple-tion method to improve participating media rendering performance,an incident light transport matrix formulation,sampling and reconstruction method,and an out-of-GPU-core rendering approach to overcome the scale limitation of GPU-based many-light ren-dering.These works overcome existing limitations and achieve up to one magnitude of speedup compared to the state-of-the-arts.The contributions include:Proposed a many-light light transport matrix sampling and reconstruction method for non-medium scene.The method employs sparse matrix sampling,visibility prediction and matrix separation techniques to significantly improve the perfor-mance of many-light rendering for non-medium scene.First,it formulates the gathering phase of many-light rendering as a light transport matrix recovery prob-lem,and cut the light transport matrix into a group of reduced matrix by slicing rows and clustering columns.Then,it use a novel visibility prediction mechanis-m that trains various visibility predictor to sparsely sample the visibility term of matrix entries.Finally,it adapts matrix separation technique to recover the sparse error in the matrix to generate the final image.In the test,this new method achieves a most 7 times of speedup over the state-of-the-art methods.Proposed a many-light light transport matrix column sampling and reconstruction method for rendering medium scene.The method employs adaptive matrix colum-n sampling and completing technique to significantly improve the performance of many-light rendering for medium scene.It explores the local coherence of light transports in participating media by formulating a novel eye ray segment and vir-tual ray light segment matrix.Then adaptively samples columns and sparsely com-plete the sampled columns.The results shows one magnitude of speedup over the state-of-the-art methods.Introduced an incident light field matrix sampling and reconstruction method.This method explores the local coherence of the incident light field among shad-ing points.It uses importance-balanced incident light field matrix reconstruction,adaptive incident directions sampling,sparse matrix completion and light field filtering techniques to improve the performance of gathering camera related infor-mation.Compared to state-of-the-art light field reconstruction and image filtering works,the new method has much better performance and quality.Introduced a GPU-based out-of-core light transport matrix organization and com-putation method for large scale out-of-core scene.The method explores the spatial coherence of light transport matrix and geometry to formulate the rendering s-cheming problem as a graphic traversal problem.It is the first many-light method that capable of processing scenes with out-of-core light and geometry data.The result demonstrates one magnitude of speedup compared to CPU-based in-core approaches.