Integrating Features, Models, and Semantics for TREC Video Retrieval

In this paper, we describe a system for automatic and interactive content-based retrieval of video that integrates features, models, and semantics. The novelty of the approach lies in the (I) semi-automatic construction of models of scenes, events, and objects from feature descriptors, and (2) integration of content-based and model-based querying in the search process. We describe several approaches for integration including iterative filtering, score aggregation, and relevance feedback searching. We describe our effort of applying the content-based retrieval system to the TREC video retrieval benchmark.