Motivated by Google's great success on text document retrieval and recent progress in semantic video understanding, researchers began to build a new generation of video retrieval systems that are able to support semantic video retrieval via keywords. Unfortunately, these systems are unable to provide satisfactory results on large-scale video collections because of several challenging problems. First, existing semantic understanding techniques are still immature, and their performance is not good enough to enable keyword-based retrieval. Second, the mismatch between visual concepts and keywords prevents any keyword-based search engines from retrieving visual concepts that are difficult to represent in language. Third, users may not have a clear idea of their needs at the beginning. Therefore, they may not be able to represent their preference with keywords.

To resolve this problem, we have proposed a systematic solution for large-scale video database exploration. The proposed system integrates achievements on semantic video understanding, information extraction, and visualization and optimizes toward a single target. The overall performance may be significantly improved via integrated optimization.

The central problem of implementing the proposed system is semantic video understanding. The system can provide useful services only if it can extract enough semantics from the video clips. To resolve this problem, we have proposed principal video shots for video content representation and a multi-class EM algorithm for effective semantic video classifier training. By integrating these two techniques, our system is able to extract sufficient video semantics. Based on semantic video understanding, we have proposed a statistical video analysis algorithm to extract value-added information in large-scale video databases. The proposed information extraction algorithm is able to extract interesting information and suppress uninteresting information. Then, we have proposed a visualization framework to visualize the extracted information. The visualization framework enables interactive and intuitive video retrieval and helps users input visual queries by clicking keywords or keyframes effectively and intuitively. In addition, we have proposed a concept-oriented subjective video skimming algorithm to help users check search results efficiently. By integrating these techniques into a system, we have achieved more effective and intuitive video exploration.