Kutulakos, Kiriakos University of Rochester $87,425.00
CAREER: Appearance-Driven Reconstruction of Three-Dimensinal Scenes in the Physical World
This is the first year funding of a four-year continuing award. This research investigates the automatic construction of geometric representations that accurately capture the three-dimensional shape, appearance, and motion of complex scenes in the physical world (objects, indoor/outdoor environments, and people). The central idea is that a reconstructed 3D'shape must be consistent with a scene's photographs in order to be a valid geometric description of that scene. In this paradigm, termed "Appearance-Driven Reconstruction," shape recovery is formulated as a geometric constraint satisfaction problem, where every input photograph is a constraint that restricts the space of all possible 3D shapes to only those shapes that reproduce that photograph. When many such photographs are available, each taken from a known position in space, they define an equivalence class of 3D shape solutions called photo-consistent shapes that simultaneously reproduce all input photographs when viewed from the photographs' viewpoints. The resulting algorithms aim to (1) compute these shapes for a hierarchy of increasingly more complex models for scene reflectance, illumination, and the image formation process, (2) provably handle scenes of arbitrary geometry and topology, (3) accept arbitrary collections of raw, unprocessed photographs as input. Educational activities will center around a new laboratory facility for Appearance-Driven Reconstruction where undergraduate and graduate students will be immersed into research as soon as possible. Course development will include (1) a new project-intensive undergraduate'course on "Visual Computing" and (2) a restructuring of the graduate computer vision course to emphasize applications in computer graphics and image-based rendering.