Spatio-temporal coherence and data reuse are important problems in digital image synthesis and processing. The existence of coherence, i.e. local data similarity, usually leads to redundancy of data and computations in virtually every stage of the pipeline. By exploiting such coherence, we can potentially reduce a large amount of unnecessary computations. This not only has the benefit of accelerating the process, but also provides opportunities to improve the result quality with the additional data that are not available otherwise.;In this thesis, we introduce techniques for spatial and temporal data reuse that benefit a number of real-time rendering and image-processing applications. For simplicity and efficiency, we explore methods that operate within the image space. Moreover, for all the applications, we seek to design parallel real-time algorithms that executes on the GPU or multi-core CPU. This may limit the class of methods that we can use, but the resulting high efficiency can benefit a much wider range of high-performance graphics applications.;For spatial data reuse, we first show how the results of interpolating sparse shading data on an image can be improved with an edge-preserving filter. We then introduce a sampling scheme that accelerates the costly computation of diffuse indirect illumination by allowing spatial data share. Moreover, in the field of image processing, we demonstrate how data in coherent regions can be reused to fix antialiased edges that are damaged by non-linear filters. For temporal data reuse, we introduce a few techniques and tools for improving the performance of data reprojection --- a fundamental operation for temporal data reuse. We then propose a technique for effectively amortizing the computation of supersampling over time. This is based on a principled analysis of the quality associated with repeated reprojection. Finally, we present an efficient frame-interpolation technique that significantly improves framerate for general real-time rendering applications.;Our proposed schemes are practical for a number of real-time rendering and image processing applications. All our methods are for interactive purposes and exhibit sufficient performance as well as result quality. We demonstrate the efficacy of our methods compared to traditional approaches. For acceleration tasks we typically observe a 4--10x speedup, and for the others we achieve new satisfactory results that are not available with previous methods. Finally, our methods are gradually gaining recognition in the industry. We propose several future directions to continue this trend of development.
展开▼
