next up previous
Next: Evidence Integration Up: Theory: Evidence and Association Previous: General Associations

Identity Inhibition

A structure seldom has more than one likely identity, unless the identities are related (e.g. a structure that is likely to be a DC-10 can also be a wide-bodied aircraft but seldom a banana). Hence, an identity should be inhibited by other unrelated identities having high plausibilities in the same context. A second source of inhibition comes from the same identity in subcontexts, to force invocation to occur only in the smallest satisfactory context. The key questions are what types provide inhibition, how to quantify the amount of inhibition and how to integrate this inhibition with the other evidence types.

Type-related inhibition is a complicated issue. Competition does not always occur even between unrelated generic types. For example, the generic type "positive-cylindrical-surface" should not compete with the generic type "elongated-surface", whereas it should compete with the generic type "negative-cylindrical-surface". The latter comparison is between two members of a set of related types that also include: "planar", "positive-ellipsoid", "negative-ellipsoid" and "hyperboloid" surfaces. All types in this set compete with each other, but not with any other types.

Inhibition results in a plausibility value like those discussed in previous sections and is then integrated with the other evidence types, as discussed below. An advantage to this method is that it still allows for alternative interpretations, as in the ambiguous duck/rabbit figure (e.g. [9]), when evidence for each is high enough.

Some constraints on the inhibition computation are:

These constraints lead to the inhibition computation:

Given:  
  a model instance of type $M$ in image context $C$
  a set {$S_i$} of all identities competing with $M$
  a set {$C_j$} of subcontexts of context $C$
  a set {$p_i$} of plausibilities for the identities $S_i$ in context $C$
  a set {$p_j$} of plausibilities for the identity $M$ in the subcontexts $C_j$
   
Then, the inhibition evidence is:
  $evd_{inh} = max(max_i(p_i),max_j(p_j))$
   

This computation gives no inhibition if no competing identities exist.

If several identities have roughly equal plausibilities, then inhibition will drive their plausibilities down, but still leave them roughly equal. A single strong identity would severely inhibit all other identities. Figure 8.12 shows the invocation network unit for computing the inhibition evidence, where the lower "max" units represent balanced trees of binary "max" units and the $p_i$ come from competing identities and the $p_j$ come from subcontexts, as described above.

Figure 8.12: Inhibition Invocation Network Fragment
\begin{figure}\epsfysize =4in
\epsfbox{FIGURES/Fig8.12.ps}\end{figure}

next up previous
Next: Evidence Integration Up: Theory: Evidence and Association Previous: General Associations
Bob Fisher 2004-02-26