A comparison of late fusion methods for object detection

In this paper, we review different classifier fusion methods and sketch the modifications which are needed for the fusion of object detection results. Four datasets are used to compare the performance of different fusion algorithms. The datasets cover different aspects of image based object detection. Hence, for each dataset a different set of sophisticated object detection methods is used. Based on the datasets and the adapted sets of base object detectors, we compare the performances of SCANN, Fuzzy Templates, Voting, AdaBoost, Random Forests, and different simple fusion methods. We report here for the first time results for CRAGORS which is a novel fusion algorithm that has been specially designed for object detector fusion. A subset of best performing fusion algorithms is identified and all methods are compared to lower and upper bounds.