Deep Anomaly Discovery from Unlabeled Videos via Normality Advantage and Self-Paced Refinement