Exploring Vision-Language Models for Weakly Supervised Video Anomaly Detection: Experiments and Analysis

Video anomaly detection (VAD) is an automated task that identifies deviations from normal behavior in video streams. Since 2023, several CLIP-based VAD methods have been proposed. These methods can be categorized into monomodal approaches, employing only visual features from CLIP, and bi-modal approaches, leveraging both visual and textual representations. However, most of these approaches have been evaluated under specific experimental settings, making it challenging to select the appropriate model for a particular application. Therefore, a comparative study that evaluates CLIP-based VAD methods under unified experimental settings is needed. This paper presents a comparative study of three CLIP-based approaches for VAD: CLIP-TSA, VadCLIP, and AnomalyCLIP. We specifically compare their performance on two publicly available datasets, XD-Violence and UCF-Crime, using the same evaluation metrics. The experimental results show that the bimodal approaches VadCLIP and AnomalyCLIP achieve similar performance, outperforming the mono-modal CLIP-TSA. Additionally, all evaluated methods perform better on XD-Violence than on UCF-Crime.