Classically, the problem of object recognition has long been formulated
as that of recognizing the instance of some object in a scene through
information acquired from an image. This has motivated such questions as
how to compute stable and effective representations for objects from
image data and how to correctly identify an unknown object from its
representation. Our interest, and the primary focus of this research, is
how such decisions can be made in the presence of uncertainty and how
one can explicitly represent and reason about the uncertainty.
We propose a new framework for parametric shape recognition based on a
probabilistic model of inverse theory first introduced by Tarantola in
[Tarantola:87]. The key result is a method for generating classifiers in
the form of conditional probability densities in order to recognize an
unknown from a set of reference models. Our procedure is
automatic. Off-line, it invokes an autonomous process to estimate
reference model parameters and their statistics. On-line, during
measurement, it combines these with a priori context-dependent
information, as well as the parameters and statistics estimated for an
unknown object, into a single description. That description, a
conditional probability density function, represents the likelihood of
correspondence between the unknown and a particular reference model.
The context of this project is three-dimensional object recognition in
which objects are represented by parametric shape descriptors and their
uncertainties. Object models are constructed through a process of *
autonomous exploration* [Whaite:94] in which a part-oriented,
articulated description of an object is inferred through successive
probes with a laser range-finding system. The set-up used to perform
experiments is a two-axis laser range-finder mounted on an inverted
PUMA-560 manipulator, and a rotary table. For any particular viewpoint,
a process of bottom-up shape analysis leads to an articulated model of
the object's shape in which each part is represented by a superellipsoid
primitive.
An experimental system has been implemented in which objects are
represented by either single or multiple parametric models (e.g.
superellipsoids, deformable solids). These experiments have indicated
that recognition performance is near perfect for cases in which complete
surface information is accessible to the algorithm, and falls off
gracefully (minimal false-positive response) when only partial
information is available. This has led to the development of a
sequential recognition strategy based on the concept of an *
informative viewpoint* in which belief measures associated with each
classification can be used as feedback for the acquisition of
further evidence as required.

T. Arbel, F.P. Ferrie

Scene and Composite Model

Mon Nov 13 10:43:02 EST 1995