Latent semantic indexing for video content modeling and analysis

In this paper we describe our method for feature extraction developed for the Video-TREC 2003 workshop. Latent Semantic Indexing (LSI) was originally introduced to efficiently index text documents by detecting synonyms and the polysemy of words. We successfully proposed an adaptation of LSI to model video content for object retrieval. Following this idea we now present an extension of our work to index and compare video shots in a large video database. The distributions of LSI features among semantic classes is then estimated to detect concepts present in video shots. K-Nearest Neighbors and Gaussian Mixture Model classifiers are implemented for this purpose. Finally, performances obtained on LSI features are compared to a direct approach based on raw features, namely color histograms and Gabor's energies.