This project advances the state of the art by utilizing geometric consistency as a mid-level visual similarity cue used to develop a visual index of a geo-located image dataset and use the attained data associations as a means to infer semantic relationship among dataset elements. The characterization of the image content in terms of the geometric and semantic elements observed in scene provides a general framework for both identifying and managing data association in large scale photo collections. The project develops such complementary data abstractions into a single framework by focusing on two main research topics: (1) Determining the geographic location where an image was taken by comparing it against a large database of geo-located urban imagery - accordingly, the challenge of balancing both search completeness and computational tractability is brought to the forefront of research efforts; and (2) Incorporating geometric structure estimates attained from large photo-collections or ground reconnaissance video/photos as a means to identify and recognize semantically meaningful elements within the reconstructed 3D-environment.
This project leverages the use geometric consistency as a visual data association primitive in order to introduce the concept of structural and semantic indexing within the development internet scale photo collection analysis systems. Moreover, by combining the complementary data abstraction levels of geometrical structure and semantic context the research team develops more efficient and robust data organization framework with applicability well beyond the studied test application of urban geo-localization.
