Concept based video retrieval is a promising search paradigm because it is fully automated and it investigates the fine grained content of a video, which is normally not captured by human annotations. Concepts are captured by so-called concept detectors. However, since these detectors do not yet show a sufficient performance, the evaluation of retrieval systems, which are built on top of the detector output, is difficult. In this report we describe a software package which generates simulated detector output for a specified performance level. Afterwards, this output can be used to execute a search run and ultimately to evaluate the performance of the proposed retrieval method, which is normally done through comparison to a baseline. The probabilistic model of the detectors are two Gaussians, one for the positive and one for the negative class. Thus, the parameters for the simulation are the two means and deviations plus the prior probability of the concept in the dataset.
|Name||CTIT Technical Report Series|
|Publisher||Centre for Telematics and Information Technology, University of Twente|
- Detector Simulation
- Video Information Retrieval