In this talk, I will address two specific topics related to the characterization of humans in a visual scene, i.e., detection and classification, and re-identification. As for detection/classification, I will report about an approach tackling this problem based on the extraction of specific powerful features, covariance matrices, which are able to finely and reliably characterize a person in terms of detection and pose classification (face orientation), even at low resolution. Concerning the re-identification issue, I will give a summary of our research, essentially based on the design of discriminant features, starting from the definition of the first descriptor (i.e., SDALF) and its evolution, up to considering multi-modal data like 3D information, and to introducing a learning method taking benefits of these features.
Finally, if time allows, a few words will be devoted to describe an interesting extension of the re-identification concept showing a recent research which exploits actual multi-modal data, like the chat style in social media systems (e.g., Skype), to characterize and re-identify a person.