BLUE compression cuts VLM calls by 53% without semantic loss
Surveillance video compressed with BLUE preserves AI understanding while slashing costs.
BLUE (Background-Limited Universal Encoder) is a novel compression approach designed for fixed-camera surveillance systems. Unlike standard codecs like H.265 that optimize for human viewing, BLUE aggressively suppresses static background redundancy while preserving foreground activity. This machine-centric approach reduces file size without degrading downstream computer vision performance, specifically for vision-language model (VLM) analytics that require semantic understanding of events and anomalies.
The team evaluated BLUE on two datasets: VIRAT (227 paired event samples from 106 clips) and CHAD (54 human-activity anomaly clips). Using a blind judging protocol, they compared raw H.265 and BLUE-compressed H.265 video through a VLM captioning pipeline. Results showed near-identical semantic inference quality: on VIRAT, mean VLM score difference was only -0.01 (0–10 scale); on CHAD, scores were 4.31 vs 4.26. Compression ratio was uncorrelated with quality loss (r = 0.004). Most striking, BLUE increased the proportion of skip-heavy P-frames from 1.4% to 53.2% on CHAD, enabling an estimated 53% reduction in VLM inference calls through simple packet-size-based skipping. This makes BLUE a practical drop-in layer for reducing both storage and compute costs in enterprise surveillance analytics.
- BLUE compression preserved VLM scores within 0.01 points on a 0–10 scale on VIRAT and 4.31 vs 4.26 on CHAD (no statistical degradation).
- BLUE increased skip-heavy P-frames from 1.4% to 53.2% on CHAD, enabling an estimated 53% reduction in VLM inference calls.
- Compression saving was uncorrelated with semantic quality loss (r = 0.004), meaning higher compression doesn't predict worse AI understanding.
Why It Matters
Enterprise surveillance systems can slash storage and AI inference costs by 53% without losing semantic accuracy.