Abstract
In [1], we proposed a two-stage retrieval framework which makes not only performance characterization but also performance optimization manageable. There, the performance optimization focused on the second stage of the retrieval framework. In this paper, we extend the method to a full two-stage performance characterization and optimization. In our retrieval framework, the user specifies a high-level concept to be searched for, the size of the image region to be covered by the concept (e.g.“Search images with 30–50% of sky”) and an optimization option (e.g. “maximum recall”, “maximum precision” or “joint maximization of precision and recall”). For the detection of each concept such as “sky”, a multitude of concept detectors exist that perform differently. In order to reach optimum retrieval performance, the detector best satisfying the user query is selected and the information of the corresponding concept detector is processed and optimized.
Besides the optimization procedure itself the paper discusses the generation of multiple detectors per semantic concept. In experiments, the advantage of joint compared to individual optimization of first and second stage is shown.
