Measurement and representation of realistic surface appearance and 3D shape. The visual appearance of real-world objects is a function of their shape and the way incident light is scattered at their surface. Our research investigates new optical devices and capture strategies that allow recovering accurate digital models of the geometry and surface appearance of physical samples. A related goal is developing representations of measured surface appearance that are compact and enable intuitive editing. This work is supported by the National Science Foundation through a CAREER grant CCF-0747220 ("The Inverse Shade Tree Framework for Material Acquisition, Analysis, and Design") and NSF grant CCF-0811493 ("A Scanning Pipeline for the Synchronous Capture of Precise 3D Shape and Surface Appearance"). We also gratefully acknowledge equipment donations by NVIDIA in support of this work.

Data caching and reuse in interactive rendering systems. The fact that animated image sequences exhibit a significant amount of spatio-temporal coherence has been heavily exploited in conventional ray-based rendering systems. Our research considers the consequences of this observation for interactive rendering systems executed on modern graphics hardware. We have developed techniques that allow reusing intermediate values generated inside a pixel shader across time (neighboring frames) and space (neighboring pixels). We have also studied several applications of these "shading caches" that include automatic optimization of procedural shaders and amortized supersampling that allows improving the quality of individual frames for a marginal increase in computational effort. This work is partially supported by an NVIDIA Professor Partnership grant ("Automated Resource Allocation for Interactive Rendering Algorithms").

Data-driven image processing. The sheer volume of visual data (images, videos, etc.) available over the Internet suggests radically new "data-driven" approaches to longstanding open problems in image processing. We are studying the particular problem of single image super-resolution in this context (this effect is often shown being used by fictionalized criminal investigators on television and in movies). The basic idea is to compute a high-resolution version of a dital image using examples harvested from the Internet that show the relationship of image data at neighboring octaves (spatial scales). Supported by the NSF through their Cluster Exploratory (CluE) program (NSF IIS-0844416 "Image Super-Resolution using Trillions of Examples") and in conjunction with Google and IBM, we are assembling a repository of 50 million on-line images stored on a large computer cluster to study the feasibility of this approach.