CaC: Hierarchical Spatiotemporal Reward Model for Video Anomaly Detection

Researchers have introduced Concentrate and Concentrate (CaC), an anomaly reward model that transitions from coarse to fine, leveraging Vision-Language Models. In the inference phase, it conducts a comprehensive temporal analysis to identify anomalous time segments, followed by detailed spatial grounding within these defined intervals, employing structured spatiotemporal Chain-of-Thought reasoning for robust conclusions. To support this model, the team created the first extensive video anomaly dataset, featuring per-frame bounding-box annotations, temporal anomaly windows, and detailed attribution labels. Their training framework is structured in three stages: the model first acquires spatial and temporal anchoring through both single- and multi-frame supervised fine-tuning, then undergoes reinforcement learning using two-turn Group Relative Policy Optimization (GRPO), enhancing video reward models beyond traditional accuracy measures.