Historically, verification has meant several different things in the context of vision. The fundamental notion is that of confirming the existence of an oriented object, but this is often reduced to merely confirming the presence of a few object features.
Typical verification methods predict image features (e.g. lines) given the model and current state of analysis, which then strengthen or weaken the hypothesis according to the presence or absence of confirming evidence (e.g. [60]). Additionally, the discrepancy between the observed and predicted position can be used to refine the position estimates [170].
The best verification work has been applied in the context of two dimensional industrial scenes, as in parts location systems (e.g. [35,109]). Object silhouettes are most often used, because they make the object contours explicit; however, edge detected grey level images also produce similar information. The most common verification feature is the edge, and usually just the straight edge is used, though small slots and holes at given distances from test features have also been used [35]. The main advantages of these features are that their shape, location and orientation are easy to predict. Prediction also allows more sensitive edge detection ([148,168]), when searching for confirming evidence.
In two dimensional scenes, overlapping parts weaken the utility of contours, because only part of each object's outline is visible, and it is also joined with those of the other objects in the pile. Since most two dimensional recognition systems are dependent on contours, this produces a serious loss of information. Yin [169] hypothesized objects based on visible corners and linear features and verified them by ensuring that all unlocated corners were within the contours of the collected mass.
Verification in three dimensional scenes has not received much attention. Some work similar to the two dimensional line verification has been done in the context of three dimensional blocks world scenes by Falk [60] and Shirai [148]. ACRONYM's [42] prediction graph informed on the observable features, their appearance and their interrelationships in the context of more complicated objects (e.g. wide-bodied airplanes). Hogg [91] verified three dimensional generalized cylinder model positions by counting oriented edge points within image boxes. The boxes were predicted using the projected outlines of generalized cylinders.
Occlusion is an even greater problem in three dimensions, as scenes have natural depth and hence objects will often self-obscure as well as obscure each other. Brooks [42] suggested that a model-based geometric reasoning vision system could predict what features will be self-obscured from a given viewpoint. In the context of the blocks world scene analysis, occlusion hypotheses were verified by detecting single TEE junctions to signal the start of occlusion (e.g. [162]) and pairs of TEEs indicated which edges should be associated (e.g. [80]).