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