This paper describes a hierarchical approach towards the extraction of highly informative summarized information of humans from video sequences. Objects of interest, such as facial features, are detected through transformation-invariant clustering of the frames, iteratively from bigger to smaller regions, and then expressed with an information-rich representation obtained by super-resolution. To guarantee the fundamental constraints under which the super-resolution process is well-behaved, we propose a Bayesian framework that integrates the uncertainties in the registration of the frames. The ultimate product of the overall process is a strip of images that describe at high resolution the dynamics of the video, switching between alternative local descriptions in response to visual changes.