Semantic indexing and known item search based on a unified model with topic transition representation


We applied a generative approach to the TRECVID 2010 Semantic Indexing (SIN) and Known-Item Search (KIS) tasks, using a probabilistic network called Hierarchical Topic Trajectory Model (HTTM). It is our newly-developed model that can integrate multiple sources of potentially associated information such as video frames and texts, as well as dynamically changing high-level pieces of information such as topics. With this model, the semantic indexing and the knownitem search tasks were dealt within a single unified framework. We show how it worked, and present some analysis for the SIN task.

2010 TREC Video Retrieval Evaluation Notebook Papers