At this point the recognition process has isolated a set of data, described its features and invoked a model as its potential identity. To claim that the object is genuinely recognized requires the pairing of model features to image data. Without these correspondences object recognition is only suggestive - like saying a collection of gears and springs is a watch. Hence, the hypothesis construction process has the goal of fully instantiating correctly invoked models, estimating object position and accounting for the appearance of the object, including occlusion. These tasks are the first stage in substantiating the existence and identity of the object, as described in this chapter.
The chapter also describes some more recent work that uses a value-propagating
network for geometric reasoning.