Cascaded Reduction and Growing of Result Sets for Combining Object Detectors

In this paper cascaded reduction and growing of result sets is introduced as a principle for combining the results of different object detectors. First, different candidate operating points are selected for each object detection algorithm. This procedure is based on the analysis of precision and recall of the individual methods. Selecting an appropriate operating point prior to fusion is important because it regulates the cardinal number of the result set. As diversity and correlation between object detectors also depend on the elements of the result sets, this and the application of set operations allow to create a final set of detected objects by including missing and excluding false detections. The approach allows both diverse and correlated detectors to contribute to the performance of the combined detector. The performance of the proposed algorithm is compared to other combining algorithms. It outperforms or competes with existing state of the art combiners for several datasets. Additionally, the results provide a significant improvement in the interpretability of the combining rules. As a unique feature of the proposed algorithm, the found operating points can be used to reconfigure the object detection algorithms to adapt their individual results to the needs of the combination procedure allowing a reduction in runtime.