In this paper we consider how shape (including patterns and objects) can be
encoded in terms of a relational learning method which simultaneously deri
ves features, their attribute ranges and the dependencies which best descri
be their specific shapes. To illustrate this approach we consider two probl
ems in the context of pattern and object recognition. First, the problem of
determining what constitutes 'features' or 'parts' of patterns? Second, th
e problem of what constitutes acceptable variations of shape in a recogniti
on process? In the former case we examine polyhedral approximations to 3D o
bjects while in the latter case we explore range-based objects defined by s
urfaces of arbitrary shape and form. The results demonstrate the robustness
and explanatory power of the approach. (C) 1999 Elsevier Science B.V. All
rights reserved.