Reproduce “CoPS: Conditional Prompt Synthesis for Zero-Shot Anomaly Detection” 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
Qiyu Chen, Zhen Qu, Wei Luo, Haiming Yao, Yunkang Cao, Yuxin Jiang, Yinan Duan, Huiyuan Luo and 2 more
The CoPS framework enhances zero-shot anomaly detection by synthesizing dynamic, visually-conditioned prompts for CLIP, utilizing explicit state token synthesis, implicit class token sampling, and spatially-aware glocal alignment. This approach improves image-level AUROC by 1.4% and pixel-level AUROC by 1.9% on average across various industrial and medical datasets.
Key insights
- Dynamically generated prompts, conditioned on visual features, can effectively overcome the limitations of static tokens and sparse labels in zero-shot anomaly detection.
- Explicitly modeling normal and anomaly states through prototypes injected into prompt tokens significantly enhances a model's ability to generalize to novel anomaly categories.
- Leveraging a VAE to implicitly sample and fuse diverse semantic class tokens enriches prompt information, reducing overfitting and improving robustness across varied categories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.