Next: Identity Inhibition Up: Theory: Evidence and Association Previous: Subclass Evidence

## General Associations

This relationship gives evidence for the presence of an object of type , given the presence of an object of arbitrary type . This is not a structure or class relationship, which have already been considered here, here, here and here. It is an "other relationships" category. An relationship of this type might be: "the presence of a desk makes the presence of a chair plausible". This allows many forms of peripheral evidence and can be thought of as a "priming" relationship.

Association is not commutative, so individual connections need to be made, if desired, for each direction. For example, the presence of a car makes the presence of a road likely, whereas there are many roads without cars. On the other hand, the evidence supplied is weak: the certain presence of the car does not necessarily mean the object under consideration is a road. Further, the strength of association will depend on the objects involved.

The previous evidence types have clearly specified contexts from which evidence came, but this type does not. Generally associated objects could be anywhere in the scene, so all nodes of the desired type give support.

Some constraints on this type of relationship are:

• The presence of an associated object increases the plausibility of the object.
• The more plausible the associated object, the more plausible the object.
• The weight of an association expresses the expectation that the desired object is present, given that the associated object is present.
• The context of the association is the whole image.

These constraints lead to the following formal definition of the general association computation:

 Given: a model instance of type in image context a set {()} of associated models (where is the strength of association) a set {} of all contexts a set {} of plausibilities of model in context Then, the association relationship evidence is:

which chooses the best evidence for each associated type, and then the best weighted association type. If no association evidence is available, this computation is not applied. The invocation network fragment for this evidence type is shown in Figure 8.11, where the "max" units represent balanced trees of binary "max" units and the come from other nodes, as described above.

Next: Identity Inhibition Up: Theory: Evidence and Association Previous: Subclass Evidence
Bob Fisher 2004-02-26