Ascender 2: Knowledge-Directed Image Understanding for Site Reconstruction
Abstract
The Ascender 2 system was designed to perform three dimensional reconstruction of cultural objects (primarily buildings) from multiple aerial images. It is based on the premise that cooperative redundant reconstruction algorithms will succeed where individual algorithms fail and that a major task for a vision system is deciding which algorithm to apply to what data (and when). Control is based on Bayesian networks and utility theory is used to compute the marginal value of information for alternative operators and to select the one with the highest return. Two reconstruction algorithms are described that, along with other techniques, form the repertoire of algorithms. One algorithm reconstructs a 3-dimensional model of the scene using the differential geometry of scene surfaces to index into a set of model surfaces. A robust surface optimization converges on the model and parameters that most closely describe the data. After the best-fit surface has been determined, an outlier analysis phase searches for substructures that are recursively processed. The second algorithm recovers geometric structure from SAR and IFSAR data. The presence of noise missing data and poorly understood radar artifacts in such images necessitates the use of robust and context-sensitive technique. The algorithm exploits knowledge about the geometric structure of buildings and how this geometry interacts with the sensor.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2001
- Accession Number
- ADA388649
Entities
People
- Allen Hanson
- Edward M. Riseman
- Howard Schultz
Organizations
- University of Massachusetts Amherst