Vries. An integrated approach to text and image retrieval – the lowlands team at TRECVID 2005

Our main focus for this year was on setting up a flexible retrieval environment rather than on evaluating novel video retrieval approaches. In this structured abstract the submitted runs are briefly described. High-level feature extraction We experimented with feature detectors based on visual information only, and compared Weibullbased and GMM-based detectors. • LL-HF-WB-VisOnly Region-based Weibull models, visual only • LL-HF-WBNWC-VisOnly Extended region-based Weibull models, visual only • LL-HF-GMMQGM-VisOnly GMM-based models, query generation variant • LL-HF-GMMDGM-VisOnly GMM-based models, document generation variant We found large differences across topics. Some models are good for one topic other for the next. Future research has to show whether a combined approach is useful. Search In the search task we focused on a seamless integration of our visual and textual retrieval system, to allow for easy multimodal querying. We use the Nexi language for querying (see Section 3.1) and Ram for specifying visual retrieval models (see Section 3.3). • M-A-1-LL-ram-text-1 manual text only run • F-A-1-LL-ram-text-2 fully automatic text only run • M-A-2-LL-ram-text-im-3 manual text + image run • M-A-2-LL-ram-text-feat-4 manual text + high-level feature run • M-A-2-LL-ram-text-im-feat-5 manual text + image + high-level feature run • F-A-2-LL-ram-text-im-6 fully automatic text + image run • F-A-1-LL-tijahpsql-text-7 fully automatic text only run We experimented with a generic retrieval approach that used collection specific information only for training the high-level feature detectors. Runs making use of textual information perform around the median, adding visual information does not influence the results. 1