Institute for Robotics and Intelligent Systems
Department of Computer Science
University of Southern California
Los Angeles, California
Goals:
- General techniques for automated map-making and photo-interpretation.
- Detection and description of buildings from monocular aerial images of general viewpoints (i.e. nadir views and oblique views).
Applications:
- Automated site model generation.
- Photo-interpretation.
- Automated map-making or cartography.
- Change detection and Virtual reality.
Difficulties:
- "Figure/Ground" problem (Low-level segmentation is fragmented).
- Lack of direct 3-D information.
Previous Work:
- Several previous systems developed (references in the paper).
- Most of them use contour tracing and the search complexity increased exponentially with image complexity.
- Some of them are working on nadir views only.
- 3-D cues in monocular images are not fully utilized.
- Our previous system used perceptual grouping and shadow techniques, but limited to nadir views.
Assumptions:
- Projection is locally weak perspective.
- Viewpoint and sun angles are known.
- Roofs are flat and rectilinear, walls are vertical.
- Shadows fall on flat ground.
Approach:
- Hypothesize roof candidates.
- Select good hypotheses based on image evidence.
- Verify hypotheses based on wall and shadow evidence.
Most Recent Results:
Modelboard Results: (09/15/95)
"Buildings Detection and Description from Monocular Aerial Images,"
Chungan Lin and Ramakant Nevatia,
ARPA Image Understanding Workshop 1996.
