Realistic computer graphics involving very large image datasets, such as lightfields, precomputed light transport, time-lapse video, measured material database, has become an increasingly important and interesting topic in recent years. While rendering systems can exploit these data to achieve greatly improved performance and visual fidelity, such datasets also present significant challenges for efficient processing, not only because of their large scale and high dimensionality, but more importantly, because of their intrinsic nonlinear structures. That is, the data naturally contain very high frequency, rapidly changing information that cannot be adequately represented using linear processing models that prevail elsewhere in computer graphics.
This project develops nonlinear processing methods for large-scale, multidimensional light transport data. These nonlinear methods adaptively select the best solutions for representation driven by the data?s intrinsic nonlinear structure, thereby significantly improving the sampling, compression, reduction, and reconstruction of such data. The project involves three steps of research. The first step collects light transport data from both simulated and measured sources, and derives a novel sampling strategy by using data-driven analysis. In the second step, the PI investigates several new methods for nonlinearly reducing and approximating light transport data into compact representations. These methods include nonlinear basis projection, clustered piecewise constant approximation, and augmented dimensionality reduction. The third step develops efficient computational methods using the compressed data representation for fast reconstruction and real-time visualization. These methods are naturally amenable to acceleration on modern graphics hardware. The resulting new approaches from this research can leverage broader applications in other fields that involve similar types of image datasets, such as geosciences, medical imaging, and forensic analysis.
