next up previous index
Next: Experiments Up: MRF-Based Matching Previous: Posterior Probability and Energy

5.2.2

Matching to Multiple Objects

The MAP configuration derived in the above is the optimal mapping from the scene to the model object under consideration. In other words, it is the optimal labeling of the scene in terms of the model object.

Suppose there are L potential model objects. Then L MAP solutions, , can be obtained after matching the scene to each of the models in turn (Fast indexing of model objects (Lamdan and Wolfson 1988) is a topic not studied in this work). However, any feature in the scene can have only one match of model feature. To resolve this, we use the following method of cost minimization.

Rewrite in (5.18) into the following form

 

where

 

and

 

are local posterior energy of order one and two, respectively. is the cost incurred by the local match given the rest of the matches. It will be used as the basis for selecting the best matched objects for i in matching to multiple model objects. The image feature i is considered to come from object if the decision incurs the least cost

 

The final label for i is feature number of object . Note, however, the MAP principle is applied to matching to a single model object not to multiple objects; the simple rule (5.22) does not maximizes the posterior since at least the partition functions are different for matching to different objects.

  
Figure 5.5: Mapping from the scene to multiple model objects. Different textures represent different structures. Bold lines represent sub-mappings. Note that the background and un-model structure are mapped to the NULL structure (the blank square). From (Li 1992c) with permission; © 1992 Elsevier.

 

Applying (5.22) to every i yields an overall mapping from the scene to the models, composed of several sub-mappings, as illustrated in Fig. 5.5. On the right, the squares with different textures represent the candidate model structures to which the scene is to be matched. Among them, the blank square represents the NULL model. On the left is the scene structure. The regions, each corresponding to a subpart of a model structure, are overlapping (un-separated). The recognition of the overlapping scene is (1) to partition the scene structure into parts such that each part is due to a single object and (2) to find correspondences between features in each part and those in the corresponding model object. The parts of the scene corresponding to the background and un-modeled objects should be mapped to the NULL , or in other words assigned the NULL label.