An approach on how to combine object recognition methods

Abstract

Object recognition methods are usually suited to specific classes of features. Nonetheless, real world objects seldom contain only one single class of features. The premise of this paper is that it would be possible to recognize more objects if several differing recognition methods were used independently on identical scene objects and their results were combined; furthermore, a combination would help to improve the quality of the recognition. First, the original scene image (which may contain overlapping objects) is preprocessed several times; every method uses its preprocessed image as input in its recognition process. Every method delivers then a quality value for every object to reflect the result of the match. In a subsequent step, each object is weighted by taking into account the given scene, the object, and the class of features peculiar to the method. This weight gives the method's suitability to characterize the particular object. The weights and quality values are then combined to produce a final list of objects sorted according to the probability that they occur in the scene. We illustrate our system combining two methods for object recognition. We then present some results and discuss the future direction of our research.