In three dimensional scenes, not all object features will be visible (e.g. some will be on the back side of an object), and a true three dimensional scene understanding program should account for this. There are three distinct cases of feature invisibility. The first case always occurs: there are features on the back side of every object and these cannot ordinarily be detected from a single viewpoint (except by using mirrors or shadows). At the same time, it is easy to predict what cannot be seen, using the estimated orientation of hypotheses to predict back-facing SURFACEs.
The next case is forward-facing self-obscured features. Here, an object feature is obscured by one or more closer surfaces from the same object. Given knowledge of the object's shape and position relative to the viewer, the relative surface positions and their visibility can be predicted.
Finally, there is structure obscured by unrelated objects. Here, the details of occlusion cannot be predicted, nor is it easy to deduce the invisible structure (though context and historical information could help - as in the top of a desk). Perhaps the best that can be done is to show that what remains is consistent with the hypothesis of obscured structure. Consequently, indirect evidence for some features must be found. This requires three actions - predicting feature visibility, finding evidence for closer structures and verifying that the available features up to the point of occlusion are consistent with the model.
After feature visibility analysis, the results are used in three ways:
These three cases of feature visibility are only applied to
individual SURFACEs, as any ASSEMBLY can be decomposed into SURFACEs.