Scholarlink

<p>In road anomaly detection, the limited information in closed - set training sets restricts models’ ability to perceive anomalies. Also, the complex traffic environment makes annotating abnormal regions in closed - set training sets quite tough. Thus, this paper proposes an open-vocabulary unsupervised road anomaly detection method (OVUR-AD) based on multi-scale cross-modal alignment. It uses language anchors from vision-language models pre-trained on large-scale datasets to tackle the closed-set problem. Then a linear multi-level cross-modal fusion strategy is proposed to strengthen the interaction between language anchors and image features. Moreover, we also design an adaptive balanced distribution module to optimize anomaly-score processing, improving the accuracy of category detection. Our method is extensively assessed on three datasets and outperforms previous unsupervised detection methods in AUPR and FPR@95. Github page:<a href="https://github.com/gitfrahts/OVUR-AD">https://github.com/gitfrahts/OVUR-AD</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

OVUR-AD: open-vocabulary unsupervised road anomaly detection via multi-scale cross-modal alignment

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Gaihua Wang,
  • Tianshun Hu,
  • Yan Guo

摘要

In road anomaly detection, the limited information in closed - set training sets restricts models’ ability to perceive anomalies. Also, the complex traffic environment makes annotating abnormal regions in closed - set training sets quite tough. Thus, this paper proposes an open-vocabulary unsupervised road anomaly detection method (OVUR-AD) based on multi-scale cross-modal alignment. It uses language anchors from vision-language models pre-trained on large-scale datasets to tackle the closed-set problem. Then a linear multi-level cross-modal fusion strategy is proposed to strengthen the interaction between language anchors and image features. Moreover, we also design an adaptive balanced distribution module to optimize anomaly-score processing, improving the accuracy of category detection. Our method is extensively assessed on three datasets and outperforms previous unsupervised detection methods in AUPR and FPR@95. Github page:https://github.com/gitfrahts/OVUR-AD