Of the Submitted and Some Additional Runs in the Semantic Indexing Task

Our experiments in TRECVID 2011 include participation in the semantic indexing and known-item search tasks. In the semantic indexing task we implemented linear and SVM-based classifiers on different low-level visual features extracted from the keyframes. In addition to the main keyframes provided by NIST, we also extracted and analysed additional frames from longer shots. The classifiers were fused using standard and weighted geometric mean. We submitted to the full task the following four runs: • PicSOM_1: Weighted geometric mean of all linear and SVM classifiers. • PicSOM_2: Weighted geometric mean of the SVM classifiers. • PicSOM_3: Geometric mean of the linear classifiers. • PicSOM_4: Geometric mean of all linear and SVM classifiers. The run PicSOM_1 obtained the highest MXIAP score of 0.1355. In the known-item search task we submitted four automatic runs: • PicSOM_1: Text search with OCR and spell check + concept detectors • PicSOM_2: Text search with OCR and spell check • PicSOM_3: Text search with lemmatisation + concept detectors • PicSOM_4: Text search with lemmatisation Our automatic runs used text search with a single video-level index containing all the ASR text plus the title, description and subjects from the meta data. We also included text detected by OCR and tried lemmatisation in some runs. In addition we used automatic selection of concepts based on matching keywords in the query text. Neither the concept detectors nor the lemmatisation managed to improve over our best run which was PicSOM_2 with a MIR score of 0.2720.