CALLM: Cascading Autoencoder and Large Language Model for Video Anomaly Detection

This paper introduces a new approach using a 3D Deep Autoencoder and a Large Visual Language Model (LVLM) to bridge the gap between video data and multi-modal models for Video Anomaly Detection. The study explores the limitations of previous architectures, particularly their lack of expertise when encountering out-of-distribution instances. By integrating an autoencoder and an LVLM in the same pipeline, this method predicts an abnormality's presence and provides a detailed explanation. Moreover, this can be achieved by employing binary classification and automatically prompting a new query. Testing reveals that the inference capability of the system offers a promising solution to the shortcomings of industrial models. However, the lack of high-quality instruction-follow video data for anomaly detection necessitates a weakly supervised method. Current limitations from the LLM domain, such as object hallucination and low physics perception, are acknowledged, highlighting the need for further research to improve model design and data quality for the video anomaly detection domain.