This paper describes ways to enhance the process of object recognition
by providing an object recognition system with the abilities to antic
ipate the occurrences of types of objects in an image and to acquire t
he information needed to form such anticipations. Two heuristics have
been identified in our study. Following the recency heuristic, one ant
icipates the reappearance in the near future of(types of) objects enco
untered in the recent past. Given an occurrence of one kind of object,
the co-occurrence heuristic sanctions anticipation of occurrences of
other types of objects that have frequently occurred with the given ki
nd of object. If these anticipations hold true, then a significant num
ber of tine regions in the image can be accounted for with a severely
pruned search space, leaving a problem significantly reduced in comple
xity. We borrow from probability theory to develop a notion of conditi
onal anticipation used by the cooccurrence heuristic, and relate the u
se of co-occurrence information to the psychology of learning and memo
ry. The proposed approach has been implemented and evaluated an the kn
owledge-based object recognition system (KOREL). KOREL automatically a
cquires models of object views and recognizes 3D objects in 2D digitiz
ed line drawing images even in the presence of modest occlusion. Exper
imental results on the enhanced recognition process are presented. (C)
1998 Pattern Recognition Society. Published by Elsevier Science Ltd.
All rights reserved.