This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).
This project provides challenging test data and benchmarks designed to advance stereo vision methods to a level of practical relevance. It aims to bridge the gap between the sophisticated but brittle methods that perform best on current benchmarks and the robust but simple methods employed in real-world applications.
The project provides new high-resolution datasets with accurate ground truth, taken with different cameras under different lighting conditions, and depicting complex indoor and outdoor scenes with non-Lambertian surfaces and outliers such as moving people, reflections, and shadows. The project explores novel algorithmic approaches for dealing with such challenges, including ways to leverage resolution, deriving color and noise models on the fly, and designing local region-growing techniques that allow deferring global optimization from the pixel level to the region level. Undergraduate students are actively involved in all components of this research.
The project has strong potential impact along several fronts. The datasets and benchmarks resulting from this work serve as catalyst for new research and enable machine learning approaches. The algorithmic contributions allow harnessing the explosion of images available online. Robust matching techniques that can handle the variety of images available on the Internet enable a host of new applications with broad impacts on the population at large, including visual localization and navigation, as well as automatic 3D reconstruction and visualization of whole cities. Finally, the project exposes undergraduates at a liberal-arts college in rural Vermont to the world of research, experimentation, and discovery.
