Abstract
The possibility to automatically index human faces in videos could lead to a wide range of applications such as automatic video content analysis, data mining, on-demand streaming, etc. Most relevant works in the literature gather full indexing of videos in real scenarios by exploiting additional media features (e.g. audio and text) that are fused with facial appearance information to make the whole frameworks accurate and robust. Anyway, there exist some application contexts where multimedia data are either not available or reliable and for which available solutions are not well suited. This paper tries to explore this challenging research path by introducing a new fully computer vision based video indexing pipeline. The system has been validated and tested in two different typical scenarios where no-multimedia data could be exploited: broadcasted political video documentaries and healthcare therapies sessions about non-verbal skills.