TREC 2003 Video Retrieval and Story Segmentation Task at NUS PRIS
Tat‐Seng Chua, Yunlong Zhao, Lekha Chaisorn, Chun-Keat Koh, Hui Yang, Huaxin Xu, Qi Chuan Tian · 2003
This paper describes the details of our systems for story segmentation task and search task of the TREC-2003 Video Track. In story segmentation task, we propose a two-level multi-modal framework. First we analyze the video at the shot level using a variety of low and high-level features, and classify the shots into pre-defined categories using a Decision Tree. Next we perform HMM analysis in order to identify news story boundaries. The two-level framework has been found to be effective in overcoming the data sparseness problem in machine learning. In the search task, we perform news video retrieval by integrating multiple intra-video features and external knowledge sources. The retrieval is performed in three stages. Stage 1 uses mainly question-ansering style text retrieval technology. It analyses the text query issued by the users and extracts relevant video stories based on ASR, and external resources like WordNet and related news articles on the web. The second stage acts as a concept filter, which eliminates the irrelevant video shots in the stories retrieved by text query system. The third stage re-ranks the retrieved shots using the image and video retrieval techniques with relevance feedback. Our system emphasizes the automated retrieve process. The experiments demonstrate the effectiveness of the story segmentation system and video retrieval system. 1.