Paper Overview
Field: Computer Vision Authors: Wenti Yin, Xiang Wang, Huaxin Zhang arXiv: 2508.03800
Abstract (Translated)
Video anomaly detection (VAD) aims to identify and temporally localize abnormal events in videos. Supervised methods learn anomaly decision boundaries from target-domain annotations but require substantial in-domain data. Existing training-free methods leverage the rich semantic knowledge and reasoning capabilities of pretrained models to interpret visual content, yet these capabilities do not directly define an anomaly decision criterion: richer anomaly descriptions better capture hazard resemblance without resolving abnormality.
To this end, we propose Contrastive Event Adjudication for training-free Video Anomaly Detection (CEAVAD), which shifts the unit of inference from isolated anomaly concepts to falsifiable event hypotheses and establishes an inference-time explanatory boundary through the interaction between competing explanations and video evidence.
Specifically, CEAVAD first constructs hazard-benign event contrasts using public safety knowledge, pairing each hazardous mechanism with a generic normal description and a mechanism-specific benign counterpart. It then determines whether the target interval better supports the hazardous explanation or its benign competitor, yielding a correctable contrastive boundary proposal for the target. Finally, CEAVAD adjudicates between the competing explanations to decide whether the hazard hypothesis withstands the video evidence, supporting both anomaly detection with temporal localization and evidence-based explanation.
Experiments on three widely used VAD benchmarks demonstrate that CEAVAD achieves state-of-the-art performance in the training-free paradigm.
Key Points
- Problem: Training-free VAD methods describe hazards richly but lack a principled anomaly decision criterion — descriptions capture "hazard resemblance" without resolving actual abnormality.
- Core idea: Move from isolated anomaly concepts to falsifiable event hypotheses, with an inference-time explanatory boundary.
- Method:
- Build hazard–benign event contrasts from public safety knowledge (each hazard paired with a generic normal description and a mechanism-specific benign counterpart).
- Compare candidate video intervals against hazardous vs. benign explanations to produce a correctable contrastive boundary proposal.
- Adjudicate between competing explanations against video evidence, enabling temporal localization and evidence-based interpretation.
- Results: State-of-the-art training-free performance on three widely used VAD benchmarks.