A new approach for computing qualitative part-based descriptions of 3D objects
from single- and multi-view range data is presented. This research is
motivated by both a theory of human image understanding
(Recognition-by-Components) and the need for qualitative recognition by an
autonomous robot in order for it to efficiently interact with its environment.

\vspace{3mm} Object descriptions are obtained in two consecutive steps: (1)
object segmentation into parts and (2) part model identification.
Segmentation is achieved by first computing the simulated electrical charge
density distribution on a tessellated triangular mesh of the object surface.
The algorithm then detects the object part boundaries where the the charge
density achieves a local minimum. The charge density distribution can
simultaneously provide an indication of the gross and fine object structures.
Parametric geons are introduced as the part models, which indicate both
qualitative shape and quantitative attribute information. Model recovery is
achieved by fitting all parametric geons to a part and then selecting the best
model based on the minimum fitting residual.  A new objective function used
for model recovery is optimised by a global optimisation technique (Very Fast
Simulated Re-Annealing).

\vspace{3mm} The advantages of this approach are demonstrated through
experimentation. By using a physical analogy to the well known transversality
principle, part segmentation does not require an assumption of surface
smoothness or the choice of a particular scale to compute local surface
features. The formulation for parametric geons provides a global shape
constraint, which ensures reliable part model recovery even when the part
shape is not an exact instance of a parametric geon. By directly comparing a
part with all candidate models, this approach explicitly verifies the shape of
the resultant part descriptions. The computed part-based descriptions are well
suited for the object recognition task carried out by an autonomous robot.