TRECVID 2004 experiments in Dublin City University

In this paper, we describe our experiments for TRECVID 2004 for the Search task. In the interactive search task, we developed two versions of a video search/browse system based on the Físchlár Digital Video System: one with text- and image-based searching (System A); the other with only image (System B). These two systems produced eight interactive runs. In addition we submitted ten fully automatic supplemental runs and two manual runs. A.1, Submitted Runs: • DCUTREC13a_{1,3,5,7} for System A, four interactive runs based on text and image evidence. • DCUTREC13b_{2,4,6,8} for System B, also four interactive runs based on image evidence alone. • DCUTV2004_9, a manual run based on filtering faces from an underlying text search engine for certain queries. • DCUTV2004_10, a manual run based on manually generated queries processed automatically. • DCU_AUTOLM{1,2,3,4,5,6,7}, seven fully automatic runs based on language models operating over ASR text transcripts and visual features. • DCUauto_{01,02,03}, three fully automatic runs based on exploring the benefits of multiple sources of text evidence and automatic query expansion. A.2, In the interactive experiment it was confirmed that text and image based retrieval outperforms an image-only system. In the fully automatic runs, DCUauto_{01,02,03}, it was found that integrating ASR, CC and OCR text into the text ranking outperforms using ASR text alone. Furthermore, applying automatic query expansion to the initial results of ASR, CC, OCR text further increases performance (MAP), though not at high rank positions. For the language model-based fully automatic runs, DCU_AUTOLM{1,2,3,4,5,6,7}, we found that interpolated language models perform marginally better than other tested language models and that combining image and textual (ASR) evidence was found to marginally increase performance (MAP) over textual models alone. For our two manual runs we found that employing a face filter disimproved MAP when compared to employing textual evidence alone and that manually generated textual queries improved MAP over fully automatic runs, though the improvement was marginal. A.3, Our conclusions from our fully automatic text based runs suggest that integrating ASR, CC and OCR text into the retrieval mechanism boost retrieval performance over ASR alone. In addition, a text-only Language Modelling approach such as DCU_AUTOLM1 will outperform our best conventional text search system. From our interactive runs we conclude that textual evidence is an important lever for locating relevant content quickly, but that image evidence, if used by experienced users can aid retrieval performance. A.4, We learned that incorporating multiple text sources improves over ASR alone and that an LM approach which integrates shot text, neighbouring shots and entire video contents provides even better retrieval performance. These findings will influence how we integrate textual evidence into future Video IR systems. It was also found that a system based on image evidence alone can perform reasonably and given good query images can aid retrieval performance.