The importance of feature visibility for the evaluation of a matching hypothesis

In the field of object recognition, it is customary to evaluate the generated matching hypotheses with different methods. Several of these methods use a weight (constants, feature statistics, etc.) to produce an improved evaluation. These weights are calculated in the training phase of the model generation and applied later to recognize an object. Usually the weights are defined independent of feature visibility. As a consequence many hypotheses are evaluated erroneously when recognizing occluded objects. To solve this problem, the weights are calculated dependent on the visibility of the corresponding features. The proposed procedure and results of using it in the recognition of several objects are presented in this paper.