Social sensing from street-level imagery: A case study in learning spatio-temporal urban mobility patterns