New Graph Models and Algorithms for Detecting Salient Structures from Cluttered Images
Abstract
This research was focused on developing new efficient algorithms to automatically detect structures of interest from cluttered images. The major accomplishments include: 1) a unified framework for edge grouping that can detect both open and closed boundaries from cluttered images, 2) a new graph model and algorithm to detect perceptually salient structures that show a certain level of boundary symmetry from cluttered images, 3) a new MCMC-based partial shape matching algorithm that can match two 2D shape contours with nonrigid shape deformation and multiple partial occlusions, 4) a new free-shape subwindow search algorithm for object localization that outperforms the state-of-the-art rectangle subwindow search algorithms, 5) two new perceptually motivated strategies for shape classification and recognition that leads to the new state-of-the-art recognition rate on the widely used MPEG7 shape data set, and 6) a new benchmark for more objective shape-correspondence performance evaluation along with a new shape-correspondence algorithm for statistical shape analysis that pre-organizes the given population of shape instances to achieve high correspondence accuracy using less CPU time.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 24, 2010
- Accession Number
- ADA519062
Entities
People
- Song Wang
Organizations
- University of South Carolina