Similarity between audiovisual contents

For retrieval purposes, large amounts of audiovisual contents are indexed before being stored. For recommendation systems as well as for a simple usage of a search engine, some information on the similarity between two contents is required in order to identify relevant contents. But the concept of similarity can be interpreted under many points of views. We will present three different and original approaches developed at the IRIT in order to associate a value to the similarity. The first one considers that the similarity consists in a similar temporal evolution of a subset of descriptors. This leads to determine the subset of features which can be used for the evaluation. The second one consists in considering the way events are organized along the temporal dimension. Similar correlations between events are likely to highlight a global similarity between two contents. The last approach consists in involving the end user in the similarity definition. For two different tasks, one may consider that two documents considered as similar for the first one, are not similar anymore for the second one. We propose an interface where the user set the position of small set of documents on the computer screen. The more the contents are similar, the closer they are expected to be one with each other. The system then tries to extract from that configuration rules which could express the similarity nature and then apply them to the whole content set in order to show it on the screen in an organized manner.