JOANNEUM RESEARCH at TRECVID 2013: Semantic Indexing and Instance Search

We participated in two tasks: semantic indexing (SIN) and instance search (INS). SIN runs We experimented with dynamic classifier selection, fusing input runs trained using SVM with RBF kernel on color and texture features. All runs runs were of type M, using parts of the IACC1 data for training, and IACC1.C as a validation set for fusion. Three runs use different methods for selecting the best classifier and determining the resulting score, the fourth uses a static classifier selection (per concept), also using IACC1.C as a validation set. • JRS1: best in terms of AP, max. AP as score • JRS2: best in terms of AP, max. score of all agreeing classifiers • JRS3: best in terms of number of correct classification, max. score of all agreeing classifiers • JRS4: static classifier selection based on performance on validation set, per concept The fused results neither outperform the best of the input classifiers, nor the static fusion. We found that the main reason for this is that our input classifiers yield better results in terms of ranking than in terms of decision boundary, thus the runs achieve better score when their MAP is determined independently rather than when the binary classification are used to select a classifier in fusion. INS runs We submitted two runs for INS: • JRS1: Matching SIFT descriptors using GPU, absolute number of matches • JRS2: Matching SIFT descriptors using GPU, relative number of matches The results of the two runs show no differences. For some of the queries, no results were found due to a low number of (reliable) interest points. For the other runs, verification was quite strict, leading to good precision for the top 10-20 ranks, but low recall.