Superclass Evidence

This relationship gives evidence for the presence of an object of class $M$, given the presence of an object of superclass S. For example, evidence for the object being a wide-bodied aircraft lends some support for the possibility of it being a DC-10. The use of superclass is not rigorous here - the notion is of a category generalization along arbitrary lines. Hence, a class may have several generalizations: an $<automobile wheel>$ may generalize to $<torus-like object>$ or $<automobile part>$. Class generalization does not require that all constraints on the generalization are satisfied by the specialization, which differentiates this evidence type from the description relationship. So, a superclass need not be a strictly visual generalization, that is, there may be different models for the object and its superclass.

Superclasses provide circumstantial, rather than direct, evidence as the presence of the $<torus-like object>$ alone does not provide serious evidence for the $<wheel>$ being present. If the object had both strong $<torus>$ and $<automobile part>$ evidence, the implication should be stronger. If the object had strong $<torus>$ and weak $<automobile part>$ evidence, then it would be less plausible for it to be a $<automobile wheel>$. Because the superclass is a generalization, its plausibility must always be at least as great as that of the object. Hence, the evidence for an object can be at most the minimum of the evidence for its superclasses.

Constraints that help specify the superclass evidence computation are:

• The presence of a superclass increases the plausibility of the object's presence.
• The more plausible the superclass, the more plausible the object.
• The plausibility of an object is less than that of a superclass.
• The context of the superclass is that of the object.

These constraints lead to the following formal definition of the superclass evidence computation:

Given:
	a model instance of class $M$ in image context $C$
	a set {$S_i$} of superclasses of $M$
	a set {$p_i$} of plausibilities of model $S_i$ in context $C$
Then, the superclass evidence is:
	$evd_{supcl} = min_i(p_i)$

If no superclass evidence is available, this computation is not applied. The portion of the network associated with this evidence is shown in Figure 8.10, where the square unit is a "min" unit (representing a balanced tree of binary "min" units) and the inputs come from the appropriate superclass nodes.

Figure 8.10: Superclass Evidence Integration Network Fragment