This project aims to advance the understanding of energy-based stereo-matching techniques. Using new stereo data sets with ground-truth disparities, the research will address the question: "What is the right energy function to minimize?" The specific goals include developing an automated system for obtaining highly accurate multi-view stereo data sets with ground truth, and designing new energy functions whose minimization yields disparities that (1) accurately model occlusion, (2) align occlusion boundaries with object boundaries using monocular cues, and (3) create reasonable disparity hypotheses in half-occluded regions. Stereo algorithms that perform well on these challenging tasks have important applications in emerging consumer-level applications, including virtual gaze correction for teleconferencing, and the creation of image-based object models for 3D visualization. The research will also focus on developing fast, approximate energy minimization techniques, including variants of graph-cut and dynamic programming methods. Undergraduate students will be actively involved in all components of this research, in particular in the data acquisition and testing stages. The intellectual merits of the project are to provide the computer vision community with high-quality, multi-view data sets with ground truth and to improve the state of the art in energy-based image-matching techniques. The broader impacts of the project include the improved applicability of computer vision methods to applications that are becoming central to society, such as telecommunication and e-commerce; and the opportunity to expose undergraduates at a liberal-arts college to the world of research, experimentation, and discovery.
