Supercomponent Associations

This relationship gives evidence for the presence of an object, given the presence of a larger object of which it may be a subpart. Though evidence typically flows from subcomponents to objects, supercomponent evidence may be available when: (1) the supercomponent has property or relationship evidence of its own, or (2) other subcomponents of the object have been found. Unfortunately, the presence of the supercomponent implies that all subcomponents are present (though not necessarily visible), but not that an image structure is any particular component. As the supercomponent evidence (during invocation) cannot discriminate between likely and unlikely subcomponents in its context, it supports all equally and thus implements a "priming" operation. This is because the computation is one of plausibility, not certainty. Weighting factors control the amount of support a structure gets. When support is given for the wrong identities, other evidence contradicts and cancels this evidence.

There are several constraints derivable from the problem that help define the evidence computation. They are:

• The presence of a supercomponent makes the presence of an object plausible.
• The more plausible the presence of the supercomponent, the more plausible the presence of the object.
• There is at most one supercomponent of any object, though there may be many candidates.
• The context of the supercomponent must contain the context of the object.

The formal definition of the supercomponent relationship computation is:

Given:
	a model instance of type $M$ in image context $C$
	a set {$S_i$} of supercomponents of $M$
	a set {$C_j$} of supercontexts of $C$, including $C$ itself and
	$p_{ij}$ is the plausibility that model $M_i$ occurs in context $C_j$
Then:
	$f_i = max_j(p_{ij})$
chooses the best evidence for a supercomponent over all supercontexts and
	$evd_{supc} = max_i(f_i)$

picks the best supercomponent evidence over all supercomponent types. The current context is also included because the supercomponent may not have been visually segmented from the object. If no supercomponent evidence is defined or observed, this computation is not applied.

The network created for this evidence type is simply a tree of binary "max" units linked as defined above.