CIF: Small: Geometric, Variational Algorithms for Radiometric-Based Shape Reconstruction

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

The investigator will develop a new class of variational geometric inverse algorithms for reconstructing dense 3D shape of objects from measurements of a scene representing an arbitrary combination of vantage points and/or resolution by employing a common mathematical framework based on generative radiometric models. Reconstruction of shape from raw sensor data is a necessary step for graphical 3D rendering and analysis in scenarios where CAD models are unavailable or impossible. Two examples to be initially explored include unfocussed camera images from different viewpoints and/or focal lengths using thin-lens modelling, and radar signals reflected from nearby objects with different antenna locations and/or wavelengths. The framework will be general enough to apply to several related sensor modalities beyond those initially investigated, ranging from infrared, acoustics, and SAR. Furthermore, having a unified model affords the freedom to generate flexible data capture and fusion schemes where not one, but multiple sets of measurements are captured under different viewpoint and sensor setting characteristics, and use the entire set of collected data to infer an estimate of reflectance and geometry that is of superior quality relative to what may be obtained in scenarios where "isolating a single cue" is not possible. For instance, with cameras, it may be impossible to fix the viewpoint while capturing images of different focus (isolating focus), or to capture perfectly sharp images because of the finite aperture of the lens (isolating viewpoint). Removing this constraint can enable applications to endoscopy, inspection of pipes and crevices, dental impressions, as well as environmental monitoring.