Part of competent object recognition is knowing where an object is - hence its three dimensional location and orientation must be determined. This information is also needed internally, as identity verification requires finding that all visible object features are correctly placed. Moreover, an estimate of an object's position enables prediction of image locations for missing features.
Invocation suggests a few model-to-data feature (e.g. surface)
correspondences to form an initial "island of stability".
From this, the reference frame
of the object relative to the viewer can be deduced
by analyzing the geometric relationships between the features.
For example, a single surface correspondence constrains the object to a
single rotational degree-of-freedom about a nominal surface normal (assuming
that the model and data normals can be paired).
A second rotation axis, whether from a second surface or from an axis of
curvature on the first surface, usually completely constrains the object's
orientation (though possibly up to a mirror image).