Including Interaction in an Automated Modelling System (*)


Stephan Heuel (**) Ramakant Nevatia
Institute for Robotics and Intelligent Systems
 University of Southern California
 Los Angeles, California 90089-0273

Ph.: (213) 740-6428, Fax: (213) 740-7877
email: heuel@ipb.uni-bonn.de, nevatia@iris.usc.edu

These WWW pages describe an approach for including operator input and guidance in an automated building extraction system. Images and MPEG movies should help the reader to understand the types of interaction (Please use Netscape Version 1.1 or higher).

A paper about this work was also be presented at the 1995 IEEE International Symposium on Computer Vision.

Contents


Short description of the approach
An approach for including interaction in an automatic building detection and description system is described. It uses intermediate results and computations of the automatic analysis to add or correct the 3-D description of a scene. The proposed method requires a minimum amount of easily obtainable information from the user. The number of interaction steps is less or equal to those of computer assisted manual systems with a final fitting step. The system is built on top of a a monocular automatic building system developed at USC and has been tested on a number of examples with good results.
Working on extraction of 3-D site models, we propose a strategy for combining the activities of an operator and a machine by taking advantage of what perceptual abilities a machine does have. Our goal is to provide a minimum amount of input to the machine and let the machine make the decisions that it can. Our approach is based on the observation that the automatic systems often work quite reliably under certain conditions and the operator should not need to do this work. Also, when automatic systems fail, they fail due to some salient difficulties. In such cases, the operator may be able to supply an indication of the difficulty or the desired result which may suffice for the machine to finish the computation.

We use a monocular automatic building system developed at USC (called USC-MABS from now on) as the underlying system. This system is designed to find rectilinear building structures from a monocular image. It uses perceptual grouping techniques to find likely candidates for building roofs and uses shadows to confirm them and to estimate their height. The system's performance is generally quite good when sufficient parts of the roof boundaries can be extracted. However, in some cases, such as when the roof is dark, the boundary of the roof with the shadow is not detected and the system fails to confirm the presence of such buildings due to lack of sufficient evidence. In such cases, a very simple guidance from the operator, just indicating that in fact a dark building is present in the vicinity suffices for the automated system to find one on its own!

Our methodology does allow for more detailed interaction with the system, in stages, and as necessary. In the worst case, the system reduces to the user having to provide all the information as is the case for most manual systems. However, we find this capability is seldom needed in our system.

The design goals for our system can be summarized as follows:

"Easy" information for the user would be qualitative information without the need of precision, like answering the question "Is in the indicated area a building visible but not detected?". The last requirement could also be stated as: the precision required by the user should be minimized.

Footnotes

(*)This work was supported in part by the Advanced Research Projects Agency of the Department of Defense and was monitored by the Army Topographic Engineering Center under Contract No. DACA76-93-C-0014.

(**) Stephan Heuel was a visiting researcher at USC from the Institute of Photogrammetry in Bonn