Chris Jaynes, Robert Collins, Yong-Qing Cheng, Xiaoguang Wang, 
     Frank Stolle, Howard Schultz, Allen Hanson, and Edward Riseman,
"Automatic Construction of Three-Dimensional Models of Buildings,"
RADIUS: Image Understanding for Imagery Intelligence,
O.Firschein and T.Strat, eds, Morgan Kaufmann, 1996, pp. 223-236.

Abstract

The University of Massachusetts Ascender system relies on a single reconstruction strategy that is tuned to extract only one class of buildings by detecting single-level, flat, rectangular rooftops. The present paper describes initial work on a more general 3D reconstruction system that deos not rely on a single type of data, and is not restricted to a small predefined class of buildings. The approach involves multiple alternative detection and reconstruction strategies, invoked by clear contextual cues, that combine a wider set of algorithms and features for generating and fusing 3D and 3D information.

Some of these strateies have a bottom-up flavor and are meant to be applied to fairly large sets of primitive elements in the absence of strong contextual constraints. Top-down strategies, on the other hand, control the combinatoric problem by reducing the size of the input set of elements through the use of appropriate spatial and model-based constraints. Currently we are examining the control and representation issues associated with an automated controller. The utility of these strategies is demonstrated by a set of case studies involving various types of roofs.

Full Paper

Click here for full paper (2602196 bytes, compressed using gzip).