Building Verification by Learning from examples
Andres Huertas
Institute of Robotics and Intelligent Systems
University of Southern California
Los Angeles, CA 90089-0273.
Introduction
- Training set consists of automatically detected and vrified buildings.
- Collect "appearance" information from training set.
- Appearance is pixel-based statistical data from multiple sources.
- Sources are EO Panchromatic, EO color, IFSAR, and HyperSpectral data.
- Analyze weak hypotheses to help verify more buildings.
- Future: Assist hypotheses formation/selection from multi source cues.
-
-
-
The Roof Profile
Sites may include buildings that have similar radiometric similatities. After
sound geometric analysis detects and describes objects of interest, weak
hypotheses (due to occlusion, for example) are studied to evaluate cues from
radiometric sources to help verify them.
Hypothesize and verify paradigms can construct large numbers of hypotheses.
Typically, a selection mechanism collects the most promissing ones, giving
a set of instances of the shape of interest. Object specific knowledge is
applied to decide which hypotheses can ve verified as instances of the
objects of interest. This knowledge includes color, material composition,
2-1/2 data and decorrelation in interferometry analysis.
Matching
The matching procedures currently under testing are based on similarity tests for
first order statistics.
Verification
This step consists of reconciling the cues obtained from various sources to help verify weak hypotheses.
- Gray-level Statistics. (See examples below).
- Color Statistics. (Under implementation).
- Interferometry Statistics. (Under implementation).
- Hyperspectral Statistics. (Under implementation).
Example Results
Experiments from Fort Hood
and Purdue Campus.
Maintained by
Andres Huertas,
huertas@usc.edu
