![[2602.19248] No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19248] No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection
Summary
The paper presents LAVIDA, a novel zero-shot video anomaly detection framework that utilizes a Multimodal Large Language Model to enhance detection capabilities without requiring real anomaly data.
Why It Matters
Video anomaly detection is crucial for security and surveillance but is hindered by the scarcity of anomalous events. This research addresses these challenges by introducing a framework that can adapt to unseen anomalies, potentially improving safety and efficiency in various applications.
Key Takeaways
- LAVIDA framework enables zero-shot video anomaly detection.
- Utilizes pseudo-anomalies to enhance model adaptability.
- Integrates MLLM for improved semantic understanding.
- Achieves state-of-the-art performance on benchmark datasets.
- Reduces computational costs through token compression.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.19248 (cs) [Submitted on 22 Feb 2026] Title:No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection Authors:Zunkai Dai, Ke Li, Jiajia Liu, Jie Yang, Yuanyuan Qiao View a PDF of the paper titled No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection, by Zunkai Dai and 4 other authors View PDF HTML (experimental) Abstract:The collection and detection of video anomaly data has long been a challenging problem due to its rare occurrence and spatio-temporal scarcity. Existing video anomaly detection (VAD) methods under perform in open-world scenarios. Key contributing factors include limited dataset diversity, and inadequate understanding of context-dependent anomalous semantics. To address these issues, i) we propose LAVIDA, an end-to-end zero-shot video anomaly detection framework. ii) LAVIDA employs an Anomaly Exposure Sampler that transforms segmented objects into pseudo-anomalies to enhance model adaptability to unseen anomaly categories. It further integrates a Multimodal Large Language Model (MLLM) to bolster semantic comprehension capabilities. Additionally, iii) we design a token compression approach based on reverse attention to handle the spatio-temporal scarcity of anomalous patterns and decrease computational cost. The training process is conducted solely on pseudo anomalies without any VAD data. Evaluations across four benchmark VAD datasets demonstrate that...
