McGlone - 4
relating them to stored 3D object models.
Choice of model type for the description phase obviously depends largely upon the
algorithm computing the description. Nodal models, incorporating surface continu
ity information, are more appropriate for algorithms using image features defined
over areas such as stereo or shading, while boundary models are more applicable to
edge-based algorithms.
An important tradeoff at this stage is the amount of prior knowledge to incorpo
rate in the model as opposed to determining the model strictly from the input
data. Assumptions about the scene characteristics such as surface continuity or
building configurations can make the solution more stable; however, the algorithm
must be able to detect when the assumptions are violated. A strategy often used
with nodal models such as deformable sheets is to begin the solution with strong
assumptions on surface smoothness and continuity, then relax these assumptions
as the solution proceeds, allowing the model to more closely fit the scene surface
[Blake and Zisserman, 1987]. Another approach is the use of multigrid methods
[Terzopoulos, 1988], which begin with a coarse grid size (small number of nodes),
and progress to finer resolutions as the solution proceeds.
A limitation of nodal models is their representation contains only discrete points,
especially when trying to represent man-made objects such as buildings. Increasing
the point spacing to improve the resolution for complicated, discontinuous objects
greatly increases the computational effort. Another problem is that a nodal repre
sentation of an object is not necessarily unique; since the nodes in two objects or in
two representations of the same object will not in general be at the same locations,
a comparison of the representations requires interpolation on the object surface in
order to generate comparable points.
Boundary representations are well-suited for edge-based algorithms; the edge and
surface attributes can change as the algorithm proceeds, however, and the imple
mentation must take this into account.
3.2. Object recognition
Recognition is the identification, within an image or images, of specific objects
within a database or the classification of an object as a member of some generic
class. This phase is often thought of as the main focus for computer vision research
[Suetens et al ., 1992], and has been responsible for much of the research in object
models, especially with the growth of model-based vision systems [Binford, 1982,
Chin and Dyer, 1986].
The most important issue for this phase is the efficiency of access to the model, es
pecially to the object features (or their imaged properties) employed for recognition,
due to the inherently combinatoric nature of the problem.
When recognizing objects as instances of some abstract class, the model must be
able to express the variation of objects within the class. If the variation is strictly
a function of size, parametric models or boundary representation models with fixed
topology and scalable dimensions are suitable. If the shape variation consists of
combinations of various primitive shapes, then some form of CSG model is appro-