UC3M AT TRECVID 2010 Semantic Indexing Task

This paper describes the experiments carried out by the UC3M team for the TRECVID 2010 high-level feature extraction task. In our previous participations in TRECVID, we have developed a modular system to facilitate the testing of several functionalities. This year we have selected a simple system configuration and we have added some elements expected to provide an additional advantage. For instance, 1) different kinds of histogram based features have been included in the system, both with a late fusion scheme and with a spatial pyramid matching configuration; 2) according to the nature of the low level features, different kernels have been used to train a set of non-linear SVMs; and, 3) to combine the outputs of the trained SVMs, two different fusion strategies have been taken into account, an average combination and a linear 1-norm SVM. Additionally, one of our runs exploits the relationships between the different categories by taking into account the provided taxonomy relationships, thus modifying the final system output for some categories. To check the advantages provided by these new system elements, we have submitted the following runs for their evaluation in TRECVID 2010: • RUN 1 (“F A UC3M 1 1”): this run is the baseline system configuration, where the local features have been included with a Spatial Pyramid Matching scheme and an average combination of the SVMs outputs has been considered. • RUN 2 (“F A UC3M 2 2”): this run has the same configuration as RUN 1, but histogram based features have been included with a late fusion configuration. • RUN 3 (“F A UC3M 3 3”): this run replicates the configuration of RUN 2, but a 1-norm SVM has been used for the late fusion stage. • RUN 4 (“F A UC3M 4 4”): this last run modifies the final output of RUN 3 for a subset of classes according to the provided taxonomy relationships. This work is part of the i3media Project (CDTI 2007 1012) and is partially funded by the Centro para el Desarrollo Tecnológico Industrial (CDTI), within the Ingenio 2010 Program. The four submitted runs have achieved average InfAP values from 0.0457 (RUN 1) to 0.0528 (RUN 3). Thus, our best run performed 24 out of all 87 submitted runs for this task.