Poster Submissions


Part1: VITALAS at the University of Sunderland

Michael Oakes, Marco Palomino and Yan Xu University of Sunderland

The EU-funded VITALAS (Video and Image Indexing and Retrieval in the Large Scale) project aims to provide a prototype system dedicated to intelligent access to professional multimedia archives, such as the INA video archive and the BELGA image archive. The VITALAS system will initially be developed as a B2B tool, but the intention is to develop technologies which can also be used by the wider public with multimedia content search engines. One main objective is to investigate novel approaches to cross-media indexing, with the aim of scaling beyond the 200 to 300 concepts handled by current state-of-the-art systems. Having produced a report on the state-of-the-art in cross media indexing, we went on to create a state-of-the art baseline system for comparison with future VITALAS developments. We chose to implement the techniques developed by the MediaMill system, and Sunderland’s task was to extract the text features associated with the TRECVID video data set. CERTH-ITI combined these features with video features extracted by themselves, and audio features extracted by the Fraunhofer Institute. CERTH-ITI presented the combined feature sets to a Support Vector Machine, to label each video shot with the relevant concepts from the 101-concept MediaMill set. Sunderland also developed a search engine designed to match text queries derived from the test data against concept descriptors derived from the training data using the TF.IDF measure.

Part 2: Image Annotation with the AdaBoost Learning Algorithm

Wei-Chao Lin, Michael Oakes and John Tait University of Sunderland

Noisy information is the primary challenge for content-based image indexing. We apply the AdaBoost learning algorithm for training data selection, which filters out noisy information and enables images to be represented by their most important values. The k-Nearest Neighbour classifier then applies a similarity measure between the training and testing data sets, so as to assign new images into their relevant keyword classes. Alternatively, AdaBoost can construct its own learning model for the classification of test set images.

Hosted by the Knowledge Media Institute