CoReVAD: A Contextual Reasoning Framework for Training-Free Video Anomaly Detection
Researchers have developed CoReVAD, a novel framework for detecting anomalies in videos without requiring task-specific training. This approach leverages a single, frozen Vision-Language Model (VLM) to generate both anomaly scores and descriptive explanations. To refine these outputs, CoReVAD incorporates a Local Response Cleaning module for vision-text alignment and a softmax-based refinement with Gaussian smoothing for temporal context. AI
IMPACT Introduces a more efficient and interpretable method for video anomaly detection, potentially reducing computational costs and improving analysis.