Reproduce “TMUAD: Enhancing Logical Capabilities in Unified Anomaly Detection Models with a Text Memory Bank” 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
Jiawei Liu, Jiahe Hou, Wei Wang, Jinsong Du, Yang Cong, Huijie Fan
TMUAD, a three-memory framework, enhances unified anomaly detection models by incorporating a class-level text memory bank to explicitly capture logical information. It achieves an image-level AUROC of 95.2 on the MVTec LOCO dataset for combined structural and logical anomalies, outperforming prior methods like UniVAD (81.9) and LogSAD (90.2), while also performing strongly on structural anomaly detection benchmarks.
Key insights
- Explicitly encoding logical information through structured text descriptions of object categories, rather than relying solely on high-dimensional image features, significantly improves the detection of logical anomalies.
- A multi-memory approach, combining class-level text, object-level image features, and patch-level image features, is necessary for comprehensively addressing both structural and logical anomalies within a unified framework.
- Object segmentation prior to text description generation is crucial for obtaining granular and explicit logical information, leading to more accurate text-based anomaly detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.