CVPR Findings2026
Logical Consistency Optimization for Few-Shot Weakly Supervised Video Anomaly Detection
Hantao Zheng, Ning Han, Yawen Zeng, Hegui Zhu, Hao Chen
Abstract
Weakly supervised video anomaly detection (WSVAD) greatly reduces annotation costs, yet its dependence on large-scale datasets remains a major bottleneck in data-scarce scenarios. Beyond simple performance drops, data scarcity causes models to learn spurious correlations, undermining the logical validity and trustworthiness of their reasoning. To address this challenge, we propose the Logical Consistency Optimization (LCO) framework, which introduces an explicit knowledge validation loop during model reasoning. LCO employs a large language model to generate human-interpretable anomaly guidelines from video descriptions and iteratively refines them through logical consistency evaluation, ensuring that the learned knowledge is both valid and logically coherent. To further enhance temporal stability, a WeightedEMA mechanism is applied to smooth frame-level predictions. Extensive experiments demonstrate that LCO achieves state-of-the-art performance, with 84.56% AUC on UCF-Crime and 73.20% AP on XD-Violence, while maintaining robustness under limited data. These results highlight that explicit logical validation is a crucial step toward building reliable and interpretable few-shot video anomaly detection systems.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection cites Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection cites Learning Transferable Visual Models From Natural Language Supervision
Builds on22
- Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
- Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma et al.NeurIPS 2022 · 22,562 citations
- Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude LearningYu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh et al.ICCV 2021 · 495 citations
- An Investigation of Why Overparameterization Exacerbates Spurious CorrelationsShiori Sagawa, Aditi Raghunathan, Pang Wei Koh, Percy LiangICML 2020 · 436 citations
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly DetectionPeng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou et al.AAAI 2024 · 220 citations
Related papers
- Text Prompt with Normality Guidance for Weakly Supervised Video Anomaly DetectionZhiwei Yang, Jing Liu, Peng WuCVPR 2024 · 55 citations
- Weakly Supervised Video Anomaly Detection with Anomaly-Connected Components and Intention ReasoningYu Wang, Shengjie ZhaoCVPR 2026 · 6 citations
- Ex-VAD: Explainable Fine-grained Video Anomaly Detection Based on Visual-Language ModelsChao Huang, Yushu Shi, Jie Wen, Wei Wang et al.ICML 2025
- Learning Event Completeness for Weakly Supervised Video Anomaly DetectionYu Wang, Shiwei ChenICML 2025
- Linguistic Relative Policy Optimization for Video Anomaly ReasoningJiaxu Leng, Jiankang Zheng, Mengjingcheng Mo, Zhanjie Wu et al.ICML 2026
