VESSI - VLM-Enhanced Support for Surveillance and Investigations
VESSI is a system that uses large vision-language models (VLMs) to analyze surveillance videos by asking questions about specific visual features through text prompts. It processes video frames by testing them against three increasingly complex prompts and scores the results using a Composite Model Utility Score (CMUS). The framework helps identify relevant visual content for crime analysis while measuring model performance metrics like coverage and hallucination risk.