Evaluation based combining of classifiers for monitoring honeybees

We report on an image classification task originated from the video observation of beehives. Biologists desire to have an automatic support to identify individual bees which are labelled with badges. Current state of the art in object detection and evaluation of classifiers is briefly reviewed. Different algorithms are evaluated. ROC- as well as precision-recall analysis show that a gradient based method performs best. We investigate, whether and how this superior method can be further improved. Therefore, a novel approach for combining classifiers based on an evaluation methodology is proposed. From the pool of classifiers those are selected which provide complementary information while operating with high precision and low recall. This approach shows superior performance compared to the stand-alone detectors. The suggested combination strategy is compared to other combining rules as well as to the oracle classifier.