Subareas: Geometric Vision, Recognition
The goal of computer vision is to make computers see the world around them. Vision algorithms are starting to impact our everyday lives more and more. They help to keep cars safely on the road, enable remote robotics operations in hazardous environments, reconstruct 3D models of cities, and organize your personal photos. The 3D Computer Vision group in the Department of Computer Science, led by Prof. Jan-Michael Frahm, conducts research in the areas of geometric computer vision and 3D reconstruction, as well as real-time and active computer vision. The Recognition group, led by Prof. Svetlana Lazebnik, develops algorithms for image understanding and object recognition.
The goal of the research being done by the 3D Computer Vision group is to develop fully automated systems for accurate and rapid 3D reconstruction of urban environments from video streams. For many applications, 3D models are more descriptive than the frames of the original video. For example, in a 3D model of a city, users can see a very large area at once, realize the spatial arrangement of the buildings at a single glance, and navigate freely to the parts that most interest them, something that would be more difficult and time-consuming using the original video.
The goal of the Recognition group is to develop algorithms to enable computers to extract semantic information from photographs. This includes understanding high-level scene categories (e.g., city, beach, forest, classroom), segmenting and identifying individual objects (cars, people, buildings, etc.), as well as identifying materials (glass, metal, wood, etc.) and surface properties (e.g., horizontal vs. vertical surfaces). The Recognition group is also developing efficient methods for learning object models from heterogeneous, loosely tagged Internet photo collections.
Another project combines the work of Profs. Frahm and Lazebnik and involves automatically creating 3D models of landmarks and geographical locations, using ordinary two-dimensional pictures available through Internet photo sharing sites like Flickr. The technique creates the models using millions of images, processing them on a single personal computer in less than a day. The method devised by the UNC researchers provides for a far richer experience and is much more scalable than current commercial systems and alternative techniques developed by other researchers.