next up previous
Next: Evaluating the Invocation Process Up: Theory: Evidence and Association Previous: Evidence Integration

Network Evaluation

The invocation network is automatically created from the model base, image surfaces and surface cluster hierarchy. One model instance node is created for each pairing of a model feature to an image feature, provided that both are compatible - image surfaces are compatible with model SURFACEs and surface clusters are compatible with model ASSEMBLYs. The model instance nodes are then connected by network fragments that compute the evidence relations defined above. The number and type of fragments used depend on the relationships defined in the model base, the surface hypotheses and the surface cluster hierarchy.

Plausibilities are calculated by value propagation in the network. Property evidence evaluations are computed for all models with property evidence requirements, and these evidence values then initiate plausibility propagation. New values are recomputed and propagated whenever the inputs to a function unit or model instance node change if the new result is more than 0.1% different from the previous result.

The ideal computation has the network computing continuously as new descriptions are computed, assuming the invocation process executes independently of the data description process. When there is enough data to cause a plausibility to go above the invocation threshold, then that model could be invoked.

However, here we used a serial implementation that computed all property evaluations initially. Then, plausibilities are propagated throughout the network, until convergence is reached. On convergence, nodes with positive plausibilities are invoked for model-directed processing (Chapter 9). Invocations are ordered from simple-to-complex to ensure that subcomponents are identified for use in making larger hypotheses.

Because an object may appear in several nested surface clusters, it makes little sense to invoke it in all of these after it has been successfully found in one. Further, a smaller surface cluster containing a few subcomponents may acquire plausibility for containing the whole object. These too should not cause invocation. The inhibition formulation partly controls this, but one active measure was also needed. After an object hypothesis is successfully verified (Chapter 10), the hypothesis is associated with the smallest surface cluster completely containing the object. Then, all surface clusters containing or contained by this cluster have their plausibility for this model set to $-1$.


next up previous
Next: Evaluating the Invocation Process Up: Theory: Evidence and Association Previous: Evidence Integration
Bob Fisher 2004-02-26