Meta-learning guided weakly supervised video anomaly detection with dual memory and temporal attention

Abstract

Weakly supervised video anomaly detection (WSAD) aims to localise anomalous events in untrimmed videos using only video-level labels. Existing multiple instance learning (MIL) methods often suffer from poor generalisation to unseen anomaly types, unstable temporal attention, and limited adaptability when only a few labelled examples are available. To address these challenges, we propose a meta-learning framework that combines Model-Agnostic Meta-Learning (MAML) with a dual-memory, transformer-based architecture. The model incorporates a dual-branch temporal attention module that captures both long-range semantic dependencies and local temporal proximity, separate memory banks for normal and abnormal prototypes with gated inhibition, metric-learning constraints, and variational latent regularisation. MAML explicitly trains the model for rapid adaptation across heterogeneous anomaly distributions, forcing it to acquire task-invariant representations rather than memorising static training statistics. Extensive experiments on two standard benchmarks yield competitive frame-level AUC of 93.60% on XD-Violence and 86.10% on UCF-Crime. One of our main contributions is the demonstration of very good metrics for zero and few-shot cross dataset transfer experiments, using only a handful of weakly labelled videos. We thus provide a useful proof of concept where MAML has been shown to learn generalized anomaly and non-anomaly representations with a transformer based architecture and a dual memory backbone. A t-SNE analysis of the memory prototypes confirms that MAML produces well-separated normal and abnormal clusters, while without meta-learning the memory banks collapse into entangled representations. The model is also shown to be computationally efficient, confirming its practical value for real-world surveillance deployment.