Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization

2026-07-01

AI Summary

This work addresses the semantic ambiguity caused by static sampling in video anomaly understanding by proposing Anom-π, the first closed-loop active video understanding framework. It formulates anomaly comprehension as a sequential decision-making task, dynamically orchestrating atomic operations—such as local backtracking, temporal expansion, and fine-grained sampling—to emulate human-like revisiting behavior. The framework introduces a trajectory-level policy alignment mechanism that balances hypothesis disambiguation against exploration cost under weak supervision. Temporal policy learning is driven by an interactive Direct Preference Optimization (iDPO) objective coupled with an Active Evidence Inquiry (AEI) utility function. Remarkably, with only 2B parameters, Anom-π achieves highly competitive performance, significantly outperforming existing large-scale models in complex scenarios.

Abstract

Video anomaly understanding (VAU) relies on sparse, context-dependent cues. However, existing passive paradigms suffer from observational aliasing, where static sampling fails to disambiguate semantically distinct events. To overcome this, we propose $Anom\text{-}π$, a closed-loop framework that reconceptualizes video understanding as an active sequential decision-making process within a dynamic environment. Inspired by human video-reviewing behavior, this framework unifies internal cognitive reasoning and strategic evidence acquisition into an interleaved policy, utilizing temporal atomic operators such as local backtracking, temporal expansion, and fine-grained sampling to endow the model with perceptual proactivity. To learn such complex interaction strategies under video-level weak supervision, we design Interactive Direct Preference Optimization (iDPO) to achieve trajectory-level policy alignment, guided by an Active Evidence Inquiry (AEI) utility that balances task success, informative evidence acquisition, and interaction cost. This approach enables the agent to learn to actively disambiguate hypotheses while suppressing redundant exploration. Extensive experiments demonstrate that our framework, with only 2B parameters, achieves highly competitive performance, significantly outperforming state-of-the-art large-scale VAU models in complex scenarios.

Problem

• video anomaly understanding
• observational aliasing
• passive paradigm
• context-dependent cues
• semantic disambiguation

Innovation

• Active Video Understanding
• Interleaved Policy Optimization
• Observational Aliasing
• Interactive Direct Preference Optimization
• Temporal Atomic Operators