Reproduce “PromptMoE: Generalizable Zero-Shot Anomaly Detection via Visually-Guided Prompt Mixtures” 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

Yuheng Shao, Lizhang Wang, Changhao Li, Peixian Chen, Qinyuan Liu

PromptMoE addresses zero-shot anomaly detection by identifying anomalous images and localizing their pixels in categories excluded from training. It combines image-guided mixtures of learned normal and abnormal prompt experts, using auxiliary annotated data and a frozen CLIP backbone. On MVTec AD, it reports 93.8% image-level AUROC, but evidence is limited to tested datasets and protocol.

Key insights

  • Robust Zero-Shot Anomaly Detection benefits from a compositional approach to prompt learning, where a basis set of 'expert prompts' can be dynamically combined to form nuanced, instance-specific textual representations.
  • The model learns distinct behaviors for 'normality' (converging to core concepts) and 'abnormality' (diverging into diverse semantic primitives), which is crucial for effective anomaly representation.
  • Expert load balancing and decoupling losses are synergistic, promoting representational diversity among experts and preventing redundancy, which is vital for enhancing generalization in compositional prompting.

Summary by alphaXiv.

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