Anomalies cannot materialize or vanish out of thin air: A hierarchical multiple instance learning with position-scale awareness for video anomaly detection
Yong Su,Yuyu Tan,Simin An,Meng Xing
DOI: https://doi.org/10.1016/j.eswa.2024.124392
IF: 8.5
2024-06-07
Expert Systems with Applications
Abstract:Weakly-supervised video anomaly detection is typically formulated as a multi-instance learning problem, assigning anomaly scores to each video snippet by learning to rank with only video-level labels. However, previous approaches that rely on snippet-level embeddings generated by task-agnostic feature extractors inevitably encounter challenges such as intra-bag similarities and frame-level entanglement. As a result, the model may exhibit significant performance degradation, particularly when the target is looming or receding. To address these issues, we present a novel weakly-supervised hierarchical position-scale awareness model that incorporates heterogeneous cross-scale correlation learning to improve detection performance. Specifically, in addition to snippet-level embedding, we employ an object detector (e.g., YOLOv5) for frame-level target detection and perform normal target clustering. By introducing a hierarchical ranking strategy, we gradually disentangle potential anomaly targets from the frame to the snippet level. Subsequently, we design a simple yet efficient position-scale awareness inference method that predicts the spatial positions and scales of looming and receding targets based on the abnormal targets with high confidence in adjacent snippets. Furthermore, we introduce heterogeneous cross-scale correlation learning to acquire the correlation between targets and snippet embeddings, enabling our model to increase attention to anomaly-related targets. Compared to previous approaches that generate only anomaly scores for each snippet, our method can locate anomalous targets, making it more suitable for practical applications. Without bells and whistles, evaluations on commonly-used VAD benchmarks: ShanghaiTech, UCSD-Ped2, Avenue, UCF-Crime and UBnormal datasets show that our method yields competitive and highly promising results compared with existing unsupervised, self-supervised, and weakly-supervised competitors. The code will be made publicly available.
computer science, artificial intelligence,engineering, electrical & electronic,operations research & management science