VITALAS at TRECVID-2008

\emph{High Level Feature Extraction runs.} \begin{enumerate} \item A\_VITALAS.CERTH.ITI\_1: Combination of early fusion and concept score fusion with feature selection. \item A\_VITALAS.CERTH.ITI\_2: Concept score fusion with feature selection. \item A\_VITALAS.CERTH.ITI\_3: Clustering within feature space and concept score fusion with feature selection. \item A\_VITALAS.CERTH.ITI\_4: Concept score fusion for selected low level features. \item a\_VITALAS.CERTH.ITI\_5: Mandatory type `a' run, concept score fusion for selected low level features. \end{enumerate} This is the first participation of VITALAS in TRECVID. In the high level feature extraction task, our submitted runs are based mainly on visual features, while one run utilizes audio information as well; the text is not used. The experiments performed aim at evaluating the effectiveness of different approaches to input processing prior to the final classification (i.e., ranking) stage. These are (i) clustering of feature vectors within the feature space, (ii) fusion of classifier output scores for other concepts and (iii) feature selection. The results indicate that (i) fusion of the classifier output of other concepts can provide valuable information, even if the original features are not discriminative, (ii) feature selection generally improves the results (especially when the original number of dimensions is high) and (iii) clustering within the feature space with small number of clusters does not seem to provide any significant additional information. \emph{Search runs.} \begin{enumerate} \item ASR ranks shots solely based on the ASR collection, \item TOP20 uses only the 20 highest scored concepts for each shot, \item SIGMA2 defines a deviation-based threshold to determine which concepts will be considered, \item DEVTOP20 combines the previous two methods TOP20 and SIGMA2, \item DEVTOP50 works as DEVTOP20 but using the top 50 concepts, \item ASR-DEVTOP20 combines ASR and concept-based ranking. \end{enumerate} Our experiments for the search task are focused on concept retrieval. We generate an artificial text collection by merging context descriptions according to the probability of each concept to occur in a given shot. To make the approach feasible, we further need to investigate techniques for pruning the dense shot concept matrix. Despite the poor overall retrieval quality, our concept search runs show a similar performance to the pure ASR run. Only the combination of ASR and concept search yields considerable improvements. Among the tested concept pruning strategies, the simple top $k$ selection works better than the deviation-based thresholding.