Diversity Measures for Majority Voting in the Spatial Domain

The classic majority voting model can be extended to the spatial domain e.g. to solve object detection problems. However, the detector algorithms cannot be considered as independent classifiers, so a good ensemble cannot be composed by simply selecting the individually most accurate members. In classic theory, diversity measures are recommended that may help to explore the dependencies among the classifiers. In this paper, we generalize the classic diversity measures for the spatial domain within a majority voting framework. We show that these measures fit better to spatial applications with a specific example on object detection on retinal images. Moreover, we show how a more efficient descriptor can be found in terms of a weighted combination of diversity measures which correlates better with the accuracy of the ensemble.