McGlone - 10
space,” containing a number of representations which can be used as needed, an
object might first be modeled as a 3-D blob, an unstructured group of range pixels.
The next step would be fitting a superquadric to the object pixels, to refine the shape,
size and location of the object. As the resolution increases, due to the vehicle coming
closer to the object, a more detailed model such as a generalized cylinder might be
used. All relevant models for a particular object are simultaneously updated to
support various types of interpretation and reasoning tasks which may require a
particular representation.
The biggest problems with this approach are maintaining and updating current ob
ject models and adding new ones as required, in a consistent manner across time, as
the views of the object change and as variable noise enters the system. Maintaining
consistency between representations of the same object could also be a problem
imagine two separate processes computing the same quantity, but from two different
representations.
5. Conclusions
As mentioned above, the optimal model type is task, object, and system specific; in
addition, models must be interchanged with other systems with different formats.
We must therefore put our efforts into carefully defining the model types used and
the transformations between them. Each model type must be understood in terms
of the information it represents explicitly and the implicit information embedded in
its primitives, the relationships between them, and in the very choice of that type
of model.
Especially when interchanging models between systems, the implicit information
must be made explicit, without system-specific assumptions or primitives. The
origin of each aspect of the model must be documented, to allow the recreation of
the object from the same original data if necessary.
If vision systems are to become more oriented toward applications and interaction
with the real-world, we must address and understand all of the characteristics of the
object models produced. Assumptions and implications of the 3-D models produced
must be rigorously defined if they are ever to become more than research paper
illustrations or flashy demos.
6. Acknowledgements
The MAPS lab site model format was a group effort, involving Steven Cochran, Jeff
Shufelt, Yuan Hsieh, and myself, with the primary design and implementation done
by Yuan Hsieh.
The research described in this paper was supported by the Advanced Research
Projects Agency (ARPA/SISTO) and the U.S. Army Topographic Engineering Cen
ter under Contracts DACA76-91-C-0014, DACA76-92-C-0036, by ARPA-ASTO
under contract DACA76-95-C-0009, and the Air Force Office of Scientific Research
(AFOSR) under Contract F49620-92-J-0318. Student research and training sup-