This paper addresses the challenge of real-world anomaly detection in surveillance videos, proposing a deep learning approach that leverages both normal and anomalous videos. The authors introduce a deep multiple instance ranking framework to learn anomaly scores for video segments, using weakly labeled training videos where only video-level labels (normal or anomalous) are available. They introduce sparsity and temporal smoothness constraints in the ranking loss function to improve anomaly localization. The paper also presents a large-scale dataset of 1900 real-world surveillance videos, capturing 13 different anomalous events and normal activities, which is used to evaluate the proposed method. Experimental results show that the proposed method outperforms state-of-the-art approaches in anomaly detection and anomalous activity recognition, highlighting the effectiveness of the proposed framework and the value of the new dataset.This paper addresses the challenge of real-world anomaly detection in surveillance videos, proposing a deep learning approach that leverages both normal and anomalous videos. The authors introduce a deep multiple instance ranking framework to learn anomaly scores for video segments, using weakly labeled training videos where only video-level labels (normal or anomalous) are available. They introduce sparsity and temporal smoothness constraints in the ranking loss function to improve anomaly localization. The paper also presents a large-scale dataset of 1900 real-world surveillance videos, capturing 13 different anomalous events and normal activities, which is used to evaluate the proposed method. Experimental results show that the proposed method outperforms state-of-the-art approaches in anomaly detection and anomalous activity recognition, highlighting the effectiveness of the proposed framework and the value of the new dataset.