Reproduce “VID-AD: A Dataset for Image-Level Logical Anomaly Detection under Vision-Induced Distraction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hiroto Nakata, Yawen Zou, Shunsuke Sakai, Shun Maeda, Chunzhi Gu, Yijin Wei, Shangce Gao, Chao Zhang

This research introduces VID-AD, a dataset for evaluating logical anomaly detection under controlled visual distractions, and proposes a text-based framework that leverages vision-language models to achieve superior robustness against low-level visual variations. The method outperformed vision-centric baselines in detecting logical inconsistencies by significant margins across diverse capture conditions.

Key insights

  • Operating in a language embedding space, decoupled from raw visual features, provides superior robustness against low-level visual distractions compared to pixel-level or patch-centric vision methods for logical anomaly detection.
  • The ability of a Vision-Language Model to generate stable and consistent logic-focused textual descriptions under varying visual conditions is crucial for the robustness of language-based anomaly detection.
  • Linguistic expressibility of logical rules in a scenario directly impacts the detection framework's reliability; scenarios allowing clear, discrete textual grounding of attributes yield more stable performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.