Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
ReactVAU addresses the prohibitive compute and memory dilution of multimodal LLMs in live surveillance by decoupling streaming video anomaly understanding into an asynchronous slow-fast system. The architecture pairs a lightweight, continuous detector utilizing Spatial Grid Folding with an Anomaly-Aware Persistent Memory that shields transient anomalous features from temporal decay. Under strict causal streaming constraints, the framework achieves competitive anomaly detection and causal reasoning across standard benchmarks while drastically cutting inference costs by awakening the heavyweight reasoning MLLM only upon event detection.
Continuous video understanding doesn't require constant multimodal LLM compute: selectively waking a heavyweight reasoner via lightweight spatial filtering delivers real-time, causal anomaly reasoning without diluting rare events.
In this paper, we propose ReactVAU, a Slow-Fast Decoupled Framework for real-time streaming Video Anomaly Understanding (VAU). Existing VAU methods rely on offline inference with global temporal sampling, which violates causality and prevents deployment in live surveillance streams. Conversely, general streaming video models satisfy causal access but dilute rare transient anomalies during memory compression and often invoke heavyweight MLLMs uniformly over long normal intervals. React VAU addresses this gap with three synergistic components: a lightweight Fast Detection Module based on Spatial Grid Folding (SGF) for continuous anomaly filtering; an Anomaly-Aware Persistent Memory (AAPM) that protects critical visual cues from temporal decay; and a heavyweight Slow Reasoning Module that remains dormant during normal streams and is awakened only by suspicious events for semantic verification and causal description. Extensive experiments on multiple benchmarks demonstrate that ReactVAU operates under strict streaming constraints while simultaneously achieving competitive performance in both anomaly detection and causal reasoning, alongside significantly enhanced computational efficiency by minimizing heavyweight MLLM invocations. Project page is available at https://huiyuiui.github.io/ReactVAU/