Uncertainty Modeling of Contextual-Connections Between Tracklets for Unconstrained Video-Based Face Recognition.