Qualitative Methods in Computer Vision
Abstract
Current object recognition systems can only recognize a limited class of objects. Objects having variable numbers of parts and only loosely constrained shapes cannot be modeled and recognized by these systems. The PI proposed the use of a data structure called the VAPOR (Variable Appearance Object Representation) model to represent objects with these kinds of variable appearances and develop a search procedure called MOSS (Model Space Search) to find instances of these models in two-dimensional image data. The VAPOR model is an idealization of the object; all instances of the model in the image are variations from ideal appearance. The variations are evaluated by the description length of the model, measured in information-theoretic bits. MOSS selects the best model for the given image data by choosing the minimal length description. It was demonstrated how the system performs in a simple domain of circles and polygons and in the complex domain of finding cloverleaf intersections in aerial images of roads.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1993
- Accession Number
- ADA264335
Entities
People
- Azriel Rosenfeld
Organizations
- University of Maryland