Comparison between SVM-Light, a search engine-based approach and the mediamill baselines for assigning concepts to video shot annotations

This paper describes work performed at the University of Sunderland as part of the EU-funded VITALAS project. Text feature vectors, extracted from the TRECVID video data set, were submitted to an SVM-light implementation of support vector machine, which aimed 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. The search engine-based approach outperformed SVM-light, but did not perform overall as well as the MediaMill baseline for text feature extraction. However, the search-engine approach is much simpler than the supervised learning approach of MediaMill, and did outperform the MediaMill baseline for 31 of the 101 concept categories.