Model SURFACEs have no substructure, so the evidence for a hypothesized model SURFACE is the associated surface image region. ASSEMBLYs are then formed by hierarchical synthesis , so previously verified subcomponents or SURFACE hypotheses are evidence for a hypothesized ASSEMBLY.
If invocation occurs, at least one subcomponent grouping must have positive plausibility, which suggests that some subcomponents are visible and are likely to have been previously recognized. (If none were invoked, then it is unlikely that the object will be invoked.) Then, verified subcomponent hypotheses become the initial evidence for the structure.
For each image structure associated with the invocation subcomponent group, all verified hypotheses of the correct types in the current image context (surface cluster) are located. Then, groups of these verified subcomponent hypotheses are combinatorially paired with the invoked model's features to create a new hypothesis, provided that:
The combinatorial matching is potentially explosive, but each image substructure generally has only a few verified hypotheses, usually arising from symmetry or ambiguity. Objects with these problems or duplicated features generate more initially consistent hypotheses, but most of these are eliminated by constraints (2) and (4).
The worst combinatorial explosion occurs with the robot lower arm, where each of the visible planar surfaces could be one of two model SURFACEs, each in two possible orientations. The curved end SURFACE has one model in two possible orientations. Altogether, there are initially 32 possible pairings using previously recognized subcomponents. Fortunately, the constraints eliminate all but two, which are indistinguishable in this scene. The upper arm is the next worst case, with eight initial pairings of which two are left after the constraints. All other initial hypothesis generations had four or fewer cases. Most invalid cases were eliminated by constraint (4).