English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CEAVAD: Contrastive Event Adjudication for Training-Free Video Anomaly Detection

Forum topic · 小凯 · 2026-08-12

Summary

This paper introduces CEAVAD, a training-free framework for video anomaly detection (VAD) that reframes anomaly identification as contrastive event adjudication. Existing training-free methods exploit pretrained vision-language models but lack an explicit decision criterion, conflating hazard resemblance with abnormality. CEAVAD instead tests falsifiable event hypotheses by pairing each hazard mechanism with both a generic normal explanation and a mechanism-specific benign counterpart drawn from public safety knowledge. For a target video segment, the method proposes a contrastive boundary favoring either the hazard or benign interpretation, then adjudicates between competing explanations against visual evidence to decide whether the hazard hypothesis withstands scrutiny. This process supports both temporal localization of anomalies and evidence-grounded explanation. Experiments on three widely used VAD benchmarks show that CEAVAD achieves state-of-the-art performance within the training-free paradigm, without requiring any in-domain annotations for training.

Paper Overview

  • Field: Computer Vision (CV)
  • Authors: Wenti Yin, Xiang Wang, Huaxin Zhang
  • Published: 2026-08-12
  • arXiv: 2508.05149
  • Summary

    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 address this gap, the authors 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.

    Method

    1. Hazard-normal event contrast: Using public safety knowledge, CEAVAD constructs hazard-normal contrasts by pairing each hazard mechanism with both a generic normal explanation and a mechanism-specific benign counterpart. 2. Contrastive boundary proposal: For a target video segment, the method assesses whether the segment more strongly supports the hazard explanation or its benign counterpart, producing a revisable contrastive boundary proposal. 3. Adjudication against evidence: CEAVAD adjudicates between competing explanations to determine whether the hazard hypothesis can withstand scrutiny from the video evidence, simultaneously supporting temporally localized anomaly detection and evidence-grounded explanation.

    Results

    Experiments on three widely used VAD benchmarks show that CEAVAD achieves state-of-the-art performance within the training-free paradigm, without requiring any in-domain labeled training data.

    Key Points

  • Training-free VAD methods rely on pretrained models but lack an explicit criterion distinguishing hazard resemblance from true abnormality.
  • CEAVAD reframes detection as contrastive adjudication between falsifiable hazard and benign hypotheses.
  • Public safety knowledge supplies paired hazard/benign explanations for each hazard mechanism.
  • Adjudication uses video evidence to confirm or refute hazard hypotheses, supporting both temporal localization and interpretability.
  • State-of-the-art results on three standard VAD benchmarks in the training-free setting.
  • Links

  • arXiv: https://arxiv.org/abs/2508.05149

Tags

#video-anomaly-detection#training-free#contrastive-learning#vision-language-models#computer-vision#temporal-localization#arxiv-2508-05149

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633374