Modeling multimedia contents through probabilistic feature signatures.

We introduce a new family of flexible feature representations for content-based multimedia retrieval: probabilistic feature signatures. While conventional feature histograms and feature signatures aggregate the multimedia objects' feature distributions exhibited in some feature space according to a partitioning, probabilistic feature signatures model these feature distributions by means of discrete or continuous probability distributions. In this way, they combine the advantages of high expressiveness and compactness, for instance through Gaussian mixture models. In this paper, we introduce the concept of probabilistic feature signatures and provide the empirical evidence of high retrieval performance when using this feature representation type. We show that probabilistic feature signatures are able to outperform conventional feature signatures.