Although the CVs within the VCR provide a sufficient subset of views for recognition and location of arbitrary views of rigid objects, the use of ``view groups'' or subsets of features is not generally effective in reducing complexity in primitive matching schemes. Other constraints, applicable alike to object or view models, such as type checking and pairwise geometry provide better results. This must be qualified; view-dependent invariants may still have a role to play in object recognition for a large set of possible models, but such invariants must be robust and ensure rapid generation of valid hypotheses.
Viewpoint modelling appears well-suited to support a variety of applications in visibility based sensor planning, including inspection of single and multiple surface features using both single and stereo sensors. Combining recognition, location, planned inspection, and fuller sensor modelling within a single framework is a useful ability of an advanced visual inspection system.
Comments to: Sarah Price at ICBL.