A new research paper introduces BLUE, a video compression method designed for surveillance systems that prioritizes semantic information for vision-language models (VLMs). Unlike traditional codecs that focus on human viewing, BLUE suppresses static background elements while retaining foreground activity. Evaluations on the VIRAT and CHAD datasets demonstrated that BLUE compression does not degrade VLM-based event and anomaly detection performance. Furthermore, BLUE significantly increases the proportion of skip-heavy frames, potentially reducing VLM inference calls by an estimated 53% and lowering overall costs. AI
IMPACT BLUE offers a path to reduce storage and inference costs for AI-powered surveillance systems without sacrificing analytical performance.
RANK_REASON Research paper introducing a novel method for video compression. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
