TRECVID 2008 Participation by MCG-ICT-CAS

For TRECVID 2008 concept detection task, we principally focus on: (1) Early fusion of texture, edge and color features TECM, abbreviation of the combined TF*IDF weights based on SIFT features, Edge Histogram, and Color Moments. (2) To improve the training efficiency and explore the knowledge between concepts or hidden sub-domains more easily and efficiently, we propose a novel method based on Latent Dirichlet Allocation (LDA): LDA-based multiple-SVM (LDASVM). We first use LDA to cluster all the keyframes into topics according to the maximum element of the topic-simplex representation vector (TRV) of each keyframe. Then, we train the annotated data in each topic for each concept. During training, unlike multi-bag SVM, we only use positive samples in current topic for the sake of retaining sample’s separability, instead of all positive samples among the whole training set, and ignore the topics with too few positive samples. While testing a keyframe for a given concept, we adopt TRV as the weight vector, instead of equal weighting strategy, to combine the SVM outputs of topic-models. (3) Introduction of Pseudo Relevance Feedback (PRF) into our concept detection system for the purpose of making re-trained models more adaptive to the test data: unlike existing PRF techniques in text and video retrieval, we propose a preliminary strategy to explore the visual features of positive training samples to improve the quality of pseudo positive samples. Experimental results demonstrate that our proposed LDASVM approach is both effective and efficient.